Download Offset Time-Emulated Architecture for Optical Burst Switching - Modelling and Performance Evaluation

Universitat Politècnica de Catalunya
Departament d’Arquitectura de Computadors
Offset Time-Emulated Architecture for
Optical Burst Switching - Modelling and
Performance Evaluation
Ph.D. thesis by
MirosÃlaw Klinkowski
Advisor:
Dr. Davide Careglio
Co-advisors:
Prof. Josep Solé i Pareta, and
Prof. Marian Marciniak
PhD Program: Arquitectura i Tecnologia de Computadors
November 2007
ACTA DE QUALIFICACIÓ DE LA TESI DOCTORAL
Reunit el tribunal integrat pels sota signants per jutjar la tesi doctoral:
Títol de la tesi: Offset Time-Emulated Architecture for Optical Burst Switching Modelling and Performance Evaluation
Autor de la tesi: Mirosław Klinkowski
Acorda atorgar la qualificació de:
No apte
Aprovat
Notable
Excel·lent
Excel·lent Cum Laude
Barcelona, …………… de/d’….................…………….. de ..........….
El President
El Secretari
.............................................
............................................
(nom i cognoms)
(nom i cognoms)
El vocal
El vocal
El vocal
.............................................
............................................
.....................................
(nom i cognoms)
(nom i cognoms)
(nom i cognoms)
To Ola and my parents
Acknowledgments to
Marian, Josep, Davide, and MichaÃl
for their measureless support, inspiration
and for guiding me through the research of this thesis
Wala, Barbara, Andrzej and MichaÃl
for encouraging me at the very beginning of this thesis
And to my family
for their love, support, and for being always with me
Contents
List of Figures
v
List of Tables
ix
Summary
xi
Resum
xv
Structure of the thesis
xix
1 Introduction
1.1 Optical Transport Networks . . . . . . . . . . . . . . . . . . . . . . .
1.2 Overview of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . .
1
2
4
PART I: Background
7
2 Optical switching architectures
2.1 Principle of operation . . . . . . . . . . . . . . . . . . . .
2.1.1 Optical circuit switching (OCS) . . . . . . . . . .
2.1.2 Optical packet switching (OPS) . . . . . . . . . .
2.1.3 Optical burst switching (OBS) . . . . . . . . . . .
2.1.4 OPS vs. OBS . . . . . . . . . . . . . . . . . . . .
2.2 Main characteristics and comparison between OCS, OBS,
2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
and OPS
. . . . . .
.
.
.
.
.
.
.
9
10
10
10
11
11
12
18
3 Optical burst switching
3.1 Overview of general OBS concepts . . . . .
3.1.1 Signalling . . . . . . . . . . . . . .
3.1.2 Architectures and functions of OBS
3.1.3 Offset time provisioning . . . . . .
3.1.4 Resources reservation . . . . . . . .
3.1.5 Contention resolution . . . . . . . .
3.1.6 Burst scheduling . . . . . . . . . .
3.1.7 Quality of service provisioning . . .
3.1.8 Network routing . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
22
22
24
27
30
31
33
34
34
i
. . . .
. . . .
nodes
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
ii
PART II: Offset Time-Emulated OBS Architecture
37
4 E-OBS architecture
4.1 Principles of E-OBS . .
4.1.1 Node architecture
4.1.2 Control operation
4.2 Characteristics of E-OBS
4.3 Summary . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
39
39
39
41
42
48
5 Modelling of E-OBS control plane
5.1 Introduction . . . . . . . . . . . . . . . . . . . . .
5.2 Modelling of control plane . . . . . . . . . . . . .
5.2.1 Control plane impacting factors . . . . . .
5.2.2 A queuing model of OBS switch controller
5.3 E-OBS controller with a single processor . . . . .
5.3.1 Queuing models . . . . . . . . . . . . . . .
5.3.2 Results . . . . . . . . . . . . . . . . . . . .
5.4 Summary . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
49
49
50
50
52
52
52
54
57
. . . . . . .
. . . . . . .
. . . . . . .
and C-OBS
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
PART III: Quality of service provisioning
59
6 QoS provisioning in OBS networks
6.1 Basic concepts of QoS in OBS networks . . . . . . . . . . . . . . . . .
6.1.1 QoS metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.2 Absolute vs. relative QoS guarantees . . . . . . . . . . . . . .
6.1.3 QoS in connection-oriented and connection-less OBS . . . . .
6.2 Categories of QoS mechanisms in OBS networks with one-way signalling
6.2.1 Control plane-related mechanisms . . . . . . . . . . . . . . . .
6.2.2 Edge-based mechanisms . . . . . . . . . . . . . . . . . . . . .
6.2.3 Core-based mechanisms . . . . . . . . . . . . . . . . . . . . .
61
62
62
62
63
63
63
64
65
7 Performance of QoS mechanisms in E-OBS
7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1.1 QoS scenario details . . . . . . . . . . . . . . .
7.1.2 Simulation scenario . . . . . . . . . . . . . . . .
7.2 Threshold selection in BD-W mechanism . . . . . . . .
7.3 Performance results . . . . . . . . . . . . . . . . . . . .
7.3.1 Burst loss probability and throughput . . . . .
7.3.2 Burst preemption vs. offset time differentiation
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
69
69
69
71
71
73
73
74
76
8 Effective burst preemption in E-OBS
8.1 Preemption rate in a buffer-less OBS node . . . . . . . . . . . . . . .
8.2 Preemption Window (PW) mechanism . . . . . . . . . . . . . . . . .
8.2.1 Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
77
79
80
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
iii
8.3
8.4
8.5
8.2.2 The length of preemptive window . . .
A single-wavelength model of PW mechanism
8.3.1 Some remarks . . . . . . . . . . . . . .
Computer simulation of PW mechanism . . .
8.4.1 Simulation scenario . . . . . . . . . . .
8.4.2 Numerical results . . . . . . . . . . . .
8.4.3 PW and FDL buffering . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
PART IV: Routing
91
9 Routing in OBS networks
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.1 Routing terminology . . . . . . . . . . . . . . . . . . .
9.1.2 Reactive and proactive burst loss reduction techniques
9.1.3 Hop-by-hop vs. explicit routing . . . . . . . . . . . . .
9.2 State of the art . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.1 Alternative routing . . . . . . . . . . . . . . . . . . . .
9.2.2 Multi-path routing . . . . . . . . . . . . . . . . . . . .
9.2.3 Single-path routing . . . . . . . . . . . . . . . . . . . .
9.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10 Isolated alternative routing strategies
10.1 Scenario under study . . . . . . . . .
10.2 Algorithms . . . . . . . . . . . . . . .
10.3 Results . . . . . . . . . . . . . . . . .
10.4 Summary . . . . . . . . . . . . . . .
81
82
85
86
86
87
89
90
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
93
. 93
. 93
. 95
. 96
. 97
. 97
. 98
. 99
. 100
for labelled E-OBS networks101
. . . . . . . . . . . . . . . . . . 101
. . . . . . . . . . . . . . . . . . 104
. . . . . . . . . . . . . . . . . . 106
. . . . . . . . . . . . . . . . . . 106
11 Optimization of multi-path routing
11.1 Routing scenario . . . . . . . . . . . .
11.2 Formulation . . . . . . . . . . . . . . .
11.2.1 Loss models of OBS network . .
11.2.2 Optimization problem . . . . .
11.3 Partial derivatives . . . . . . . . . . . .
11.3.1 NR-LL model . . . . . . . . . .
11.3.2 R-LL model . . . . . . . . . . .
11.3.3 Remarks . . . . . . . . . . . . .
11.4 Implementation issues . . . . . . . . .
11.5 Performance . . . . . . . . . . . . . . .
11.5.1 Comparison of routing schemes
11.6 Summary . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
111
111
112
112
115
116
116
118
118
119
120
120
123
iv
PART V: Conclusions and future works
125
12 Conclusions and future works
127
Acronyms
131
Bibliography
133
A Related publications
A.1 Papers . . . . . . . . . . . . . . . .
A.2 Papers under submission . . . . . .
A.3 Contribution to European projects
A.4 Other publications . . . . . . . . .
149
149
152
152
152
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
List of Figures
1.1
1.2
Generic architecture of optical DWDM network. . . . . . . . . . . . .
The trend of migration in optical networking. (courtesy of [Pro05]) .
2
3
2.1
2.2
2.3
2.4
Optical circuit switching network. . . . . . . . . . . . . . . . . . . . .
Optical packet switching network. . . . . . . . . . . . . . . . . . . . .
Optical burst switching network. . . . . . . . . . . . . . . . . . . . .
Overview over key parameters determining circuit/burst/packet granularity and required switching technology. . . . . . . . . . . . . . . .
10
11
12
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
Signalling protocols in OBS networks. .
OBS ingress edge node. . . . . . . . . .
Burst control packet format. . . . . . .
OBS core switching node. . . . . . . .
Offset time provisioning architectures. .
Resources reservation schemes. . . . . .
Contention resolution mechanisms. . .
Burst scheduling algorithms. . . . . . .
.
.
.
.
.
.
.
.
23
25
26
27
29
31
32
34
4.1
4.2
40
4.3
4.4
4.5
4.6
4.7
General E-OBS core node architecture. . . . . . . . . . . . . . . . . .
Fiber delay coils; a) a single component, b) a part of an OBS test-bed.
(courtesy of Newport Corp., and OITDA) . . . . . . . . . . . . . . .
Time dependencies in E-OBS. . . . . . . . . . . . . . . . . . . . . . .
Unfairness in conventional OBS. . . . . . . . . . . . . . . . . . . . . .
Burst loss probability vs. remaining hops number. . . . . . . . . . . .
Burst loss probability vs. offered traffic load. . . . . . . . . . . . . . .
JIT resources reservation in a) C-OBS, and b) E-OBS. . . . . . . . .
40
41
42
43
44
45
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
Data and control networks of OBS. . . . . . . . . . . . .
Exemplary controller architectures. . . . . . . . . . . . .
General OBS control-plane queuing model. . . . . . . . .
Queuing models: a) M/M/1 with reneging, b) M/D/1/K.
Intensity of control packet arrival. . . . . . . . . . . . . .
Loss probability of control packets. . . . . . . . . . . . .
Delay budget vs. normalized mean burst duration. . . . .
Delay budget vs. average burst length. . . . . . . . . . .
.
.
.
.
.
.
.
.
50
51
52
53
54
55
56
56
6.1
Categories of QoS mechanisms in OBS networks. . . . . . . . . . . . .
64
v
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
vi
6.2
Selected QoS mechanisms in OBS networks. . . . . . . . . . . . . . .
65
7.1
7.2
Evaluated QoS network scenario. . . . . . . . . . . . . . . . . . . . .
Performance of BD-W mechanism (c = 8), a) HP class BLP, b) LP class
BLP, c) throughput, d) threshold value guaranteeing BLPHP ≤ 10−4 .
Performance of QoS mechanism vs. link dimensioning (ρ = 0.8, αHP =
30%), a) HP class BLP, b) LP class BLP, c) overall BLP, d) effective
data throughput. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Burst loss probabilities vs. HP class relative load in OTD and BP
mechanisms (ρ = 0.8, c = 8), a) HP class, b) LP class. . . . . . . . . .
Effective throughout vs. HP class relative load in OTD and BP mechanisms, with overall traffic load: a) 0.5, b) 0.8. . . . . . . . . . . . . .
70
7.3
7.4
7.5
Percentage of additional signalling necessary to release preempted burst
at each node, with HP class load: a) 30%, b) 50%. . . . . . . . . . . .
8.2 Principles of the preemption window mechanism. . . . . . . . . . . .
8.3 The length of preemptive window in PW mechanism. . . . . . . . . .
8.4 Markov chain representing a single-wavelength model of PW mechanism.
8.5 (a) Successful preemption and (b) preemption fail cases; the processing
times are neglected for simplicity, T is the duration of the Preemption
Window, lLP and lHP are the durations of the LP and HP bursts respectively, t is the arrival time of the HP control packet. . . . . . . .
8.6 Simulation vs. modeling results (ρ = 0.8, α = 0.3, µ = 2). . . . . . . .
8.7 Burst blocking probability as a function of T comparing Gaussian and
Exponential traffic models (α = 30%, ρ = 0.8, W = 16). . . . . . . . .
8.8 Burst blocking probability as a function of T and of W (α = 30%,
ρ = 0.8, Gaussian traffic model). . . . . . . . . . . . . . . . . . . . . .
8.9 Burst blocking probability as a function of ρ comparing Gaussian and
Exponential traffic models and different α (T = 10µs and W = 32). .
8.10 Burst blocking probability as a function of T (normalized to 1/µ) for
different W and FDL buffer size (α = 25%). . . . . . . . . . . . . . .
72
74
75
75
8.1
9.1
79
80
81
82
83
86
87
88
88
89
Routing algorithms. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
94
10.1 Network topologies; a) SIMPLE, b) NSFNET, and c) EON. . . . . .
10.2 Isolated alternative routing algorithms: a) PER, and b) BPR. . . . .
10.3 Burst loss probability in PER, a) SIMPLE (32λ), b) NSFNET (32λ),
and c) EON (64λ). . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.4 Burst loss probability in BPR, a) SIMPLE (32λ), b) NSFNET (32λ),
and c) EON (64λ). . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5 Amount of bursts experiencing given number of hops in BPR, a) SIMPLE (32λ, ρ = 0.8), b) NSFNET (32λ, ρ = 0.8), and c) NSFNET
(32λ, ρ = 0.5). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
103
105
107
108
109
11.1 Example of OBS network with multi-path source-based routing; x1 and
x2 are the splitting factors and x1 + x2 = 1. . . . . . . . . . . . . . . 112
vii
11.2 Link load models: a) non-reduced OBS, b) reduced OBS, and c) reduced CS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.3 Accuracy of NR-LL model, in SIMPLE, NSFNET, and EON topologies, with 8, 32, and 64 wavelengths per link, respectively. . . . . . .
11.4 Burst loss probability in OR, a) SIMPLE (32λ), b) NSFNET (32λ),
and c) EON (64λ). . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.5 Comparison of optimized multipath source routing with isolated alternative routing strategies, a) SIMPLE (32λ), b) NSFNET (32λ), and
c) EON (64λ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
114
115
121
122
viii
List of Tables
2.1
Advantages, drawbacks and foreseen for future implementations . . .
19
4.1
Advantages and drawbacks of offset-time provisioning architectures .
48
5.1
Performance of queuing models. . . . . . . . . . . . . . . . . . . . . .
53
6.1
Characteristics of QoS mechanisms in OBS networks with one-way
signalling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
67
9.1
9.2
9.3
Classification of literature on alternative routing in OBS networks. . .
Classification of literature on adaptive multi-path routing in OBS networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Classification of literature on adaptive single-path routing in OBS networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
97
98
99
10.1 Network topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
ix
x
Summary
The evolution of the transport networking is driven by continuously increasing traffic
demand due to the introduction of broadband Internet access and new end-user business applications as well as the continuing paradigm shift from voice to data services.
These trends have emerged at the same time as the advance in optical technologies
which has enabled the development of high-capacity transmission systems. The role
of optics in communication networks is often limited to the realization of transmission
functions, however, the next-generation networks will perform either some or all the
switching and control functions in the optical domain. As a result the optical transport networks will provide a global transport infrastructure for legacy and new IP
services (IP over DWDM).Optical burst switching (OBS) technology is a promising
solution for reducing the gap between transmission and switching speeds in future
networks. It offers both flexibility and efficiency through the exploitation of statistical multiplexing in optical domain.
This thesis presents the analysis, modelling, and evaluation of the optical burst
switching network with an emulated offset-time provisioning (E-OBS). E-OBS defines
an OBS network architecture to transport and switch optical data bursts in a core
network. On the contrary to a conventional offset-time provisioning OBS (C-OBS)
architecture, where a transmission offset time is introduced in the edge node, in an
E-OBS network the offset time is provided in the core node by means of an additional
fibre delay element. The architecture is motivated by several drawbacks inherent to
C-OBS architectures.
Due to the limitations in optical processing and queuing, OBS networks need a
special treatment so that they could solve problems typical of data-centric networks.
Contention resolution in optical domain together with quality of service (QoS) provisioning for quality demanding services are, among other things, the main designing
issues when developing OBS networks. Another important aspect is routing problem,
which concerns effective balancing of traffic load so that to reduce burst congestion
at overloaded links. Accounting for these requirements, the design objectives for the
E-OBS architecture are (i) feasibility of offset-time provisioning, (ii) an overall high
quality of service, and (iii) reduction of network congestion. These objectives are
achieved by combining selected concepts and strategies, together with appropriate
system design as well as network traffic engineering.
xi
xii
Part I of this thesis provides the background information for the design of E-OBS.
First, in order to motivate the application of optical burst switching concept, Chapter
2 reviews the general characteristics, requirements, and trends of switching architectures considered for next generation optical networks. In particular, it introduces
optical circuit switching, optical packet switching, and optical burst switching architectures. Then Chapter 3 discusses the features of optical burst switching as well as
the-state-of-research solutions that are considered for OBS networks.
Part II provides the discussion on functionality and feasibility of an E-OBS architecture. Chapter 4 introduces principles of E-OBS operation as well as it demonstrates
that C-OBS possesses many drawbacks that can be easily avoided in E-OBS. Some of
the discussed issues are the problem of unfairness in resources reservation, difficulty
with alternative routing, complexity of resources reservation algorithms, efficiency of
burst scheduling, and complexity in QoS provisioning. The feasibility of offset time
provisioning with the assistance of a fibre delay element is investigated in Chapter 5.
First, several factors that have impact on the control plane operation are discussed.
In a poorly-engineered network the congestion in control plane may delay excessively
the processing of control packets in an electronic core-node controller and as a result lead to the loss of data bursts. In order to approach this outcome effectively
two queueing models, which represent the operation of an exemplary OBS node controller, are introduced. The analyzed models allow to expose some relations which
exist between key OBS system parameters. Using obtained results the feasibility of
E-OBS operation with commercially available fibre delay elements is confirmed.
Part III addressed the problem of QoS provisioning. Chapter 6 discusses some
basic concepts of QoS as well as it presents the state of the art mechanisms dealing
with QoS in OBS. The discussion is supported by a qualitative comparison of the
mechanisms. Chapter 7 complements the study with a quantitative comparison of
the performance of selected, most addressed in the literature, QoS mechanisms in an
E-OBS scenario. As an outcome a burst preemption mechanism, which is characterized by the highest overall performance, is qualified for operating in E-OBS. Since
the preemptive mechanism may produce the overbooking of resources in any OBS
network, Chapter 8 addresses this issue. Particularly, it proposes a preemption window mechanism to solve the problem. Then it provides an analytical model which
legitimates correctness of the solution.
Part IV concerns the routing problem in OBS networks. Chapter 9 introduces
general routing terminology, and based on these terms it classifies different routing
strategies that have been proposed in the OBS literature. As a continuation, Chapter
10 studies several adaptive routing algorithms, which are based on isolated alternative
routing, designed for labelled E-OBS networks. The routing objective is to help the
contention resolution algorithms in the reduction of burst losses by balancing the link
loads and avoiding excessive congestion. Since the proposed algorithms might have
some difficulty with the congestion reduction, Chapter 11 provides another solution,
which is based on a non-linear optimization framework. In the scope of this proposal,
xiii
two optimization models for multi-path source routing are formulated and solved.
Then some related implementation issues are discussed.
Concluding, E-OBS is shown to be a feasible OBS network architecture of profitable functionality, to support efficiently the QoS provisioning, and to be able to
operate with different routing strategies and effectively reduce the network congestion.
It should be emphasized that the work is a part of the research activities performed
by the Advanced Broadband Communication Center (CCABA) at The Department
of Computer Architecture of the Technical University of Catalonia. In particular, the
work is carried on within five relevant research projects: the IST-2002-506760 NOBEL
(Next Generation Optical Network for Broadband in Europe) project, the IST-2002001933 E-PHOTON/ONE (Optical Networks: Towards Bandwidth Manageability
and Cost Efficiency) project, the COST 291 (Towards Digital Optical Networks)
action and the COST 293 (Graphs and Algorithms in Communication Networks)
action, all off them founded by the European Commission, and the CATARO project
founded by of the Spanish Ministry of Education and Science.
xiv
Resum
L’evolució de les xarxes publiques de transport de dades destaca per el continu augment de la demanda de tràfic a la que estan sotmeses. La causa és la imparable
popularització d’Internet i del seu ús per a tot tipus d’aplicacions, i en concret de les
de banda ampla i interactives (des de veu sobre IP fins a nous serveis de dades com
Storage Area Networks i Grid Computing), tant per part dels usuaris residencials
com des de negocis.
Aquesta tendència ha esdevingut paral·lela a l’avenç de les tecnologies òptiques
que han permès el desenvolupament dels sistemes de transmissió d’alta capacitat que
alhora la fan possible. De totes maneres, el paper de la tecnologia òptica en les xarxes
de telecomunicacions es limita encara molt sovint només a la transmissió/recepció,
tanmateix, en la pròxima generació de xarxes aquesta tecnologia prendrà un paper
cada vegada més rellevant també en la commutació i les funcions de control i gestió
de la xarxa. Com a resultat d’això les xarxes de transport òptiques proporcionaran
una infrastructura de transport global per a tota mena de serveis basats en IP (IP sobre DWDM). En aquest sentit, les xarxes de commutació de ràfegues òptiques (OBS:
Optical Bursts Switching) són una solució extraordinàriament prometedora tant per
la flexibilitat que ofereixen com per el seu alt rendiment fruit de l’explotació de la
multiplexació estadı́stica en el domini òptic.
Aquesta tesi presenta l’anàlisi, modelització i avaluació de les xarxes de commutació de ràfegues òptiques basades en l’emulació del temps de compensació (emulated offset time: E-OBS). El concepte d’E-OBS defineix una arquitectura de xarxa
OBS per al transportar i commutar ràfegues òptiques en una xarxa troncal en la que,
al contrari de l’arquitectura convencional (C-OBS) en la que el temps de compensació
s’introdueix des dels nodes d’accés, el temps de compensació s’introdueix en cadascun
dels nodes de la xarxa per mitjà d’un retardador de fibra addicional. L’arquitectura
E-OBS permet superar algunes de les desavantatges inherents a arquitectures C-OBS,
però la seva gran virtut és la compatibilitat amb les xarxes de commutació de circuits
òptics (OCS: Optical Circuit Switching) actuals i les futures xarxes de commutació
de paquets òptics (OPS: Optical Packet Switching), de manera que les xarxes OBS
basades en una arquitectura E-OBS) poden facilitat enormement la transició de unes
a les altres.
xv
xvi
A ala vista dels principals requeriments de disseny de les xarxes OBS, que són la
resolució de contencions en el domini òptic, la provisió de qualitat de servei (QoS) i
l’òptim encaminament de les ràfegues per tal de minimitzar la congestió de la xarxa,
. en aquesta tesi es proposa un disseny de l’arquitectura E-OBS basada en (i) un
mètode viable per a la provisió del temps de compensació, (ii) una qualitat alta
global de servei, i (iii) un mecanisme d’encaminament que minimitzi congestió de
xarxa.
La primera part d’aquesta tesi proporciona la informació documental necessària
per al disseny d’E-OBS. Primer es revisen les caracterı́stiques generals, requisits, i
tendències de les principals arquitectures de commutació òptica. En particular, ens
centrem en la commutació de circuits òptica, la commutació de paquets òptica i la
commutació de ràfegues òptiques; per a, mes endavant centrar-nos en l’estat de l’art
de les xarxes de commutació de ràfegues.
La segona part se centra en l’estudi de la funcionalitat i viabilitat de l’arquitectura
E-OBS. S’introdueixen els principis d’operació d’E-OBS i s’identifiquen els principals
esculls que presenten les arquitectures C-OBS i que deixen de ser-ho en una arquitectura E-OBS. Alguns d’aquests esculls són la dificultat d’utilitzar un algorisme
d’encaminament amb rutes alternatives, la complexitat dels algoritmes de reserva de
recursos i la seva falta d’equitat, la complexitat en la provisió de la QoS, etc. En
aquesta segona part es constata que l’arquitectura E-OBS redueix la complexitat dels
de reserva de recursos i es verifica la viabilitat d’operació i de funcionament de la
provisió del tremps de compensació en aquesta arquitectura a partir de figures de
comportament obtingudes amb retardadors de fibra comercialment disponibles.
La tercera part encara el problema de la provisió de la QoS. Primer s’hi revisen els
conceptes bàsics de QoS aixı́ com els mecanismes de tractament de la QoS per a xarxes
OBS fent-ne una comparació qualitativa i de rendiment de tots ells. Com a resultat
s’obté que el mecanisme que presenta un millor comportament és el d’avortament de
la transmissió de les ràfegues de més baixa prioritat quan aquestes col·lisionen amb
una de prioritat més alta (es l’anomenat Burst Preemption mechanism), el qual en
alguns casos presenta un problema de senyalització innecessària. Aquesta tercera part
es conclou amb la proposta d’un mecanisme de finestra a afegir al esquema de Burst
Preemption que només funciona sobre una arquitectura E-OBS i que soluciona aquest
problema.
En la quarta part s’afronta el problema de l’encaminament en xarxes OBS. Es
comença per revisar la terminologia general d’encaminament, i en base a aquesta
terminologia es classifiquen les principals estratègies d’encaminament que s’han proposat en la literatura per a ser utilitzades en xarxes OBS. A continuació, s’estudia
el comportament dels algoritmes d’encaminament adaptatius, els aı̈llats amb rutes
alternatives i els multicamı́ distribuı̈ts, sobre xarxes E-OBS. A la vista dels resultats
no massa satisfactoris que s’obtenen, es planteja una solució alternativa que es basa
en model d’optimització no lineal. Es formulen i resolen dos models d’optimització
xvii
per als algoritmes encaminament de font multicamı́ que redueixen notablement la
congestió en les xarxes OBS.
Finalment, aquesta tesi conclou que l’arquitectura E-OBS és factible, que és més
eficient que la C-OBS, que proveeix eficaçment QoS, i que és capaç d’operar amb
diverses estratègies d’encaminament i de reduir eficaçment la congestió de xarxa.
xviii
xix
Structure of the thesis
Environment
Optical Burst Switching (OBS): is a photonic network technology which overcomes the
wavelength switching inefficiency by a proper exploitation of the statistical multiplexing
in the optical layer. On the contrary to optical packet switching, OBS handles large
data bursts aggregated from the client packets in order to reduce the processing and
switching requirements. Moreover, a burst control packet is transmitted in a dedicated
control channel and delivered with some offset time prior to the data burst.
In this dissertation
The emulation of offset-times by means of a fiber span applied in the core node is a
possible way of provisioning the offset time in OBS networks. On the contrary to a
conventional OBS, where the offset time is introduced in the edge node by delaying the
transmission of a burst with respect to its control packet, the offset time-Emulated OBS
(E-OBS) has not been studied extensively. In this dissertation we show the advantages
of E-OBS architecture and address several issues related to network modelling, QoS
provisioning and routing over this architecture.
Offset-time provisioning
Related work:
• C-OBS (Conventional OBS) and E-OBS architectures
Contributions:
• Functional analysis of offset time provisioning methods
Modelling of E-OBS
Related work:
• State of the art on C-OBS node architecture/network dimensioning/desing issues
Contributions:
• Modelling of the OBS control plane operation
• Estimation of feasible E-OBS system parameters
xx
QoS provisioning
Related work:
• Existing QoS mechanisms for OBS networks
• Burst preemption mechanisms with signalling overhead, in detail
Contributions:
• Qualitative and quantitative comparison of the most referenced QoS mechanism
• Estimation of the signalling overhead in a burst preemption mechanism
• Proposal of a new preemption-based mechanism without signalling overhead
Routing
Related work:
• Alternative, multi-path, single-path routing mechanisms
• LP formulations for routing problem
Contributions:
• Performance evaluation of isolated alternative routing
• Non-linear optimization of multi-path source routing
Chapter 1
Introduction
The telecommunication networks are experiencing a continuous increase in demand
for transmission capacity. This trend is strictly related with the exponential growth of
the Internet. The evolution of Internet technologies is accompanied by development
of miscellaneous user and network applications; peer-to-peer (P2P) data/multimedia
file exchange, video broadcasting, grid services are among the most bandwidthdemanding applications. Simultaneously, we can observe a big progress in the deployment of broadband access network technologies (e.g., ADSL, WLAN, or FTTH)
which place immense traffic on the metro and core transport networks [KMM+ 05].
As a consequence, the next generation networks should offer high transmission and
switching capacities in order to cope with such increasing traffic demands.
Another consequence of the expansion of the Internet is the continuing paradigm
shift from voice to data services. This trend is followed by the migration of telecommunication industry from the voice-optimized to the IP-centric networks. In traditional voice-centric networks (such as SONET/SDH [G.700]) the connection-oriented
circuit-switched operation is both powerful and efficient. In particular, the traffic
characteristic of a superposition of several different constant-bit-rate connections is
the total sum and is still constant bit rate. In sharp contrast to, the traffic characteristics in data-centric networks are typically characterized as being very bursty.
Also, the packets, which are the basic data transport units, are often variable in the
length. Additionally, while a typical voice call is quite predictable in its duration
(i.e. the length of a conversation) the duration of the data transmission session in
the Internet can vary by orders of magnitude and is often characterized by heavytailed distributions. All this results in so-called self-similar nature of Internet traffic
[LTWW94] which has a direct impact on network dimensioning; in particular, buffer
sizing is crucial.
Electronic IP router architectures are not scalable enough and will suffer from
technological limitations when trying to reach the multi-terabit throughput range
[PPP06]. Indeed when the bit rates increase the density of integration of electronic
circuits increases as well. This may produce unwanted (parasitic) capacitances and
impedances at small dimensions. Another problem is power consumption and heat
dissipation at high integration level. Finally, electronic technology has its speed
limitations.
1
Chapter 1. Introduction
2
Client networks
Core node
- IP, SONET/SDH, ATM, ...
- All-optical switching function
- Processing of control information
(mainly electronically)
Optical network
Edge node
- Adaptation function
WDM links
- Simultaneous data transmission
on different wavelengths
Figure 1.1: Generic architecture of optical DWDM network.
The observed trends have emerged at the same time with advances in optical
transmission technologies which led to wavelength division multiplexing (WDM) systems (e.g., see [Cav00][Kar02]). The dense WDM (DWDM), which is an extension
of WDM, is able to accommodate up to hundreds of wavelengths; hence providing
huge transmission capacities. Accelerated development of optical networking has
been possible due to feasibility of integrated optics for both passive and active optical
components [Chi01].
1.1
Optical Transport Networks
A generic architecture of optical transport network (see Figure 1.1) consists of source
and destination edge nodes and intermediate core nodes that are connected by WDM
links. Client networks (IP, ATM, SONET/SDH, etc.) are connected to the edge nodes
where there is some adaptation function placed, which is responsible for conversion
of data signal from its input form to an optical format used in the optical network.
This function can perform for instance a simple conversion of wavelength, if the
client network is an optical network, or more sophisticated data aggregation/assembly
operation. The data in its optical form is transmitted through WDM links towards
core nodes; the WDM technology allows to transmit several data signals on different
wavelengths at the same time. The core nodes are responsible for processing of control
information and all-optical switching of data signals. In most cases the processing of
control information is performed electronically. When the data reaches the destination
edge node it is converted back to its client signal format.
Chapter 1. Introduction
time
3
All-optical
Static l NG- Dyn. l BS over dynamic l
SDH (OCS) WR-OBS APSON ORION G.709FS OBS OPS
Field
deployment
2015
Standards,
product
status
Research,
lab
status
2010
2005
From semi-static to dynamically re-configurable optical networks
technology
Figure 1.2: The trend of migration in optical networking. (courtesy of [Pro05])
Looking for a transport architecture of future optical Internet a few network
architectures have been proposed (e.g., see [LES00][PPP06]). These optical network architectures differ with respect to the degree of optical transparency and
the flexibility of optical interconnection [GDW03]. Taking into account the current status of optical technologies the short-term solutions will take advantage of
a less flexible circuit-switching model in optical circuit-switched (OCS) networks
[CGK92][VKM+ 01]. Nonetheless, the fact that the Internet is a connection-less
packet-based network is the main driver to develop, in a longer perspective, datacentric optical transport networks. In this context two other switching architectures
are considered by the research community, namely, optical packet switching (OPS)
and optical burst switching (OBS) [XPR01]. In the perspective of network optimization the implementation of packet/burst switching techniques directly in the transport
network will bring more statistical sharing of physical resources and will reduce the
connection costs.
The expected migration of switching functions from electronic to optics will be
gradual and will take place in several phases (see Figure 1.2) [KM01]. Nowadays,
transport networks (SDH/ATM/IP) are based mainly on static point-to-point and
ring connections. While the role of optics in these networks is limited mainly to the
realization of transmission functions the next-generation networks are expected to
perform dynamic optical switching as well. The scenario of such a migration anticipates different mid-term optical network solutions like for instance next-generation
SDH (NG-SDH), wavelength-routed optical burst switching (WR-OBS), automatic
switched optical networks (ASON), the Ontario research and innovation optical net-
Chapter 1. Introduction
4
work concept (ORION), or G.709 approach [Pro05]. Although, the main functions
related with node and network control will remain in the electronic domain, still, it
is very likely that some simple control operation will be also performed by means of
optical processing [DHL+ 03].
Considering these trends the optical burst switching technology can be seen as
a promising solution. In particular, it reduces the gap between transmission and
switching speeds in future networks as well as it offers both flexibility and efficiency
through the exploitation of statistical multiplexing in optical domain. It should be
underlined that great research activity has been undertaken in the field of OBS over
the past years. A number of contributions have been focused on the functionality
and performance of OBS node and network architectures. On the other hand, the
limitations of optical technologies bring some operational difficulties in conventional
OBS architectures. The burst contention problem, unfairness in access to transmission resources, complexity of control, the inherent lack of quality of service (QoS)
guarantees, poor performance of shortest-path routing, are some of the examples.
Taking into account these trends and requirements, this thesis analyzes, models,
and evaluates an offset time-emulated OBS architecture (E-OBS). E-OBS facilitates
both node and network operation with the assistance of local offset time provisioning
in core switching nodes. In addition, E-OBS supports efficient QoS provisioning as
well as it eliminates some constraints in OBS routing management.
The particular contributions of this thesis include:
• functional analysis of offset time provisioning methods,
• modelling of the OBS control plane operation,
• estimation of feasible E-OBS system parameters,
• qualitative and quantitative comparison of the most referenced QoS mechanisms,
• estimation of the signalling overhead in a burst preemption mechanism,
• proposal of a new preemption-based mechanism without signalling overhead,
• performance evaluation of isolated alternative routing,
• non-linear optimization of multi-path source routing.
1.2
Overview of the Thesis
This thesis is structured into the background part on optical switching architectures
(Chapter 2) and optical burst switching (Chapter 3) as well as the parts on the
analysis and modelling (Chapters 4-5), QoS provisioning (Chapters 6-8), and routing
(Chapters 9-11) in E-OBS.
Chapter 1. Introduction
5
• Chapter 2 introduces OCS, OPS, and OBS architectures. It also reviews their
general characteristics, requirements, and trends.
• Chapter 3 overviews the main features of optical burst switching technology.
• Chapter 4 carries out a discussion on C-OBS and E-OBS architectures. It
demonstrates that C-OBS possesses many drawbacks that can be easily avoided
in E-OBS. Some of the issues discussed in this chapter are: the problem of unfairness in resources reservation, difficulty with alternative routing, complexity
of resources reservation algorithms, efficiency of burst scheduling, and complexity in QoS provisioning.
• Chapter 5 studies feasible system parameters for E-OBS operation. The factors
that have impact on the control plane operation are discussed. In order to
approach the problem of excessive processing delays in an OBS switch controller,
two queueing models are studied. The obtained results demonstrates some
relations which exist between key OBS system parameters. Also the feasibility
of commercially available fibre delay elements for E-OBS operation is verified.
• Chapter 6 discusses some basic concepts of QoS as well as it reviews the state
of the art mechanisms dealing with QoS in OBS. The discussion is supported
by a qualitative comparison of the mechanisms.
• Chapter 7 complements the study with a quantitative comparison of the burst
loss probability performance of selected QoS mechanisms. As an outcome a
burst preemption mechanism is qualified to be the most efficient QoS method
for E-OBS networks.
• Chapter 8 introduces a preemption window mechanism which prevents from
the overbooking of resources in preemption-based OBS networks. An analytical
model of the mechanism is derived.
• Chapter 9 introduces general routing terminology. Then it classifies different
routing strategies considered for OBS networks.
• Chapter 10 studies two isolated alternative routing algorithms designed for labelled E-OBS networks. The objective is to help the node in the contention
resolution problem, and thus, to reduce burst losses.
• Chapter 11 provides another routing solution which is based on non-linear routing optimization. In the scope of this proposal, two optimization models for
multi-path source routing are formulated and solved. Also some implementation issues are discussed.
• Chapter 12 summarizes this thesis and presents an outlook to future work.
PART I
Background
Chapter 2
Optical switching architectures
Optical networks adopt several switching models that have been deployed successfully
in electronic networks.
In traditional voice-communication networks, which apply a circuit-switching model of operation, the communication between end users is achieved with the assistance
of dedicated, and established in the connection-setup phase, circuit (or channel) connections. Such dedicated circuits cannot be used by other users during the connection
is active even if no communication is taking place at the moment. Thanks to relatively
simple maintenance of circuit-connections the adaptation of circuit-switching model
to optical circuit switching (OCS) networks was considered from the very beginning
[CGK92].
On the contrary to the voice-oriented networks, the data-centric networks apply
a packet-switching paradigm. In such networks the end-to-end communication is
achieved thanks to the transmission of small packets that carry portioned information.
The packets are routed between network nodes over data links that are shared with
other traffic. In order to preserve from the collision of packets a packet buffering
is applied. As a consequence of the success of electronic data packet networks, an
immense effort has took place, beginning from the mid of 1990s, in the research on
optical packet switching (OPS) network technology [Chi95][CVG+ 98].
A burst switching, applied initially in the block transfer of asynchronous transfer
mode (ATM) networks [I.300], is another switching paradigm adopted to optical networks. A burst switched network is a packet-like network where each switching node
extracts control information from incoming packet header in advance in order to establish and maintain the appropriate switch connection for the duration of incoming
burst of data packets. Optical burst switching (OBS) architectures were proposed in
the late 1990s [QY99][Tur99] with the objective to overcome both the low flexibility
of OCS architectures and technological limitations of OPS architectures.
9
Chapter 2. Optical Switching Architectures
10
OXC
l-switching
controller
Client networks
OCS network
Switching times:
>>ms
E/O
l conv.
Connect. duration:
minutes/hours/...
Lightpaths
WDM links
Figure 2.1: Optical circuit switching network.
2.1
2.1.1
Principle of operation
Optical circuit switching (OCS)
The operation of optical circuit switching networks is connection-oriented. In particular, the transmission of data from a source node to a destination node is realized on
pre-established paths called the light-paths (see Figure 2.1). OCS switching nodes
are referred to as optical cross-connects (OXC). An OXC is responsible for all-optical
switching of data carried on an input wavelength (usually denoted as λ) in the input
port to an output wavelength in the output port. In OCS networks the smallest
switching entity, called later the granularity, is a wavelength.
Typical connection durations are expected to be even as low as some seconds and
the connection setup and release can be performed during some ms.
2.1.2
Optical packet switching (OPS)
In optical packet switching networks data packets are statistically multiplexed in
optical domain and link wavelength resources are shared between packets belonging
to different connections (see Figure 2.2). Control information is carried in packet
headers and it is extracted in each OPS node (router). An entirely optical OPS
router is supposed to process this control information in an optical way (e.g., see
[DHL+ 03]). Nevertheless, due to still immature all-optical processing the header is
usually converted to its electrical form and processed in an electronic node controller.
The controller configures a switching matrix so that the packet payload is switched
and buffered in an all-optical way.
The transmission times of typical IP packet at 10Gbps range from tens of ns to
approx. 1µs and further decreases at higher bit-rates.
Chapter 2. Optical Switching Architectures
11
OPS core router
Router
Controller
Client networks
OPS network
Switching times:
ns
Assembler
Packet size:
tens B ¸ kB
In-band signalling
WDM links
Packets
Control
headers
Figure 2.2: Optical packet switching network.
2.1.3
Optical burst switching (OBS)
In the optical burst switching the wavelength resources are shared between different
connections, similar to OPS. At the edge of an OBS network, the packets coming
from legacy networks (e.g., IP, ATM networks) are aggregated into large optical data
bursts which are further transmitted and switched in the network (see Figure 2.3).
Each burst has assigned a control packet. The burst control packet and its data
payload are transmitted separately on dedicated wavelengths. The control packet is
delivered to a core switching node with some offset time prior to the burst payload.
In such a way an electronic controller of the core node has time both to process the
control information and to setup the switching matrix for the incoming burst. The
burst crosses the configured nodes remaining all the way in optical domain.
The duration of typical burst, which aggregates a group of packets, can last from
some µs to several hundreds of ms.
2.1.4
OPS vs. OBS
During the past years the definition of OBS and OPS has become less clear because
of the large number of proposals claiming either name. Both burst switching and
packet switching models provide sub-channel granularity by employing asynchronous
time division multiplexing. In case switching is performed all-optically and data stays
in the optical domain until the destination edge node the concepts can be referred to
as optical burst switching and optical packet switching.
Following characteristics either individually or in combination can be regarded
defining for OBS in contrast to OPS:
Chapter 2. Optical Switching Architectures
12
OBS core node
Reservation
Manager
Client networks
OBS network
Switching times:
ns¸ms
Assembler
Burst size: kB¸MB
Out-of-band signalling
WDM links
Control
channels
Offset
Data
channels
Figure 2.3: Optical burst switching network.
1. client layer data is aggregated and assembled into larger variable length optical
data units in edge nodes,
2. control information is signalled out-of-band, processed electronically in all core
nodes and used to set up the switch matrix before the data bursts arrive.
2.2
Main characteristics and comparison between
OCS, OBS, and OPS
In order to recognize the main characteristics of optical switching architectures we
lead a detailed discussion on several control and data plane-related issues.
Hardware requirements
There is a significant difference in the switching time requirements of each optical
switching architecture (see Figure 2.4). This together with the various switching
granularity (circuits/bursts/packets, i.e., switching speeds of ms for burst switching with end-to-end setup, µs for burst switching with one-pass reservation, and ns
for packet switching) is reflected in different requirements for both applied opticalswitching components [PPP03] and electronic node controllers [VGPMGH+ 07].
The switching function in OCS nodes can be achieved with commercially available optical switch technologies such as for instance micro-electro-mechanical systems
(MEMS) [CLP02]. In fact, relatively long MEMS switching times are sufficient for
low-dynamic OCS operation. On the other hand, a dynamic character of optical burst
and packet switching requires a fast-switching operation. N s-scale switching times
Chapter 2. Optical Switching Architectures
transmission duration
(granularity)
13
Burst
Packet
Dynamic Circuit
PLZT
SOAs
switching time
MEMS
TWCs + AWG
1
100
10
nanosec.
1
100
10
microsec.
1
100
10
milisec.
1
100
10
second
Figure 2.4: Overview over key parameters determining circuit/burst/packet granularity and required switching technology.
can be achieved with commercially available arrayed waveguide grating (AWG) and
tuneable wavelength converter (TWC) technologies or the semiconductor optical amplifier (SOA) technology (e.g., see [CDB+ 03][TZ99]). The lead lanthanum zirconate
titanate (PLZT) technology [NTL+ 05] with sub-µs switching times can be a good
choice for optical burst switching.
The huge amount of quasi-simultaneously arriving OPS packets may intensify the
congestion in control plane. As a result the OPS node controller, which cannot sustain
the control traffic load, starts dropping optical packets. One method to alleviate this
effect is to introduce high-speed network processors [VGPMGH+ 07]. Although OBS
architectures with the aggregated burst transmission reduce the congestion of control
processing, still, they have to be designed carefully in order to prevent from data
losses [BD07].
The complexity of hardware is also related with the node architecture, functionality of optical components and equipment dimensions. The highest demands of the
above-mentioned issues become apparent for the OPS [BIPe99] which is still looking
forward to more advanced compact optical elements. OBS architectures with processing offsets and long data burst transmissions have relatively moderate technological
requirements comparing to OPS.
Further we can distinguish the following technological requirements for optical
packet and burst switching:
• Suitable switch fabric technology with low loss, low crosstalk, low polarization
dependence and low power consumption.
• Burst mode receivers that allow for fast synchronization of clock and adjustment
of decision threshold, synchronization/adaptation speed requirements depending on and adapted to switching speed.
• Optical regenerators including wavelength conversion which are tuneable (arbitrary or tuneable/selectable input wavelength and/or tuneable output wavelength) with tuning speed equal to switching speed.
Chapter 2. Optical Switching Architectures
14
• Tuneable transmitters with tuning speed equal to the switching speed at high
output power and high side mode suppression ratio.
• Efficient fibre delay line structures for optical buffering and synchronization
especially for OPS.
Processing complexity
The processing complexity of control algorithms at the node level is much higher at
OBS/OPS than in OCS. In the formers, specific algorithms for contention resolution,
data assembling (at the edge node) and QoS provisioning have to be implemented to
be able to fully exploit the potential benefits of the inherent statistical multiplexing
principle.
Essentially, the complexity at the network level corresponds to the complexity of
routing algorithms or a path calculation process. In the case of OCS, a calculation of
the light-path has to occur prior to its establishment in the network, although once
established, little maintenance is required at intermediate core nodes. In a connectionless OPS/OBS environment a path has to be calculated at each node and for each
packet or burst, respectively. Since each packet/burst potentially has TE and QoS
related attributes it needs to be taken into account during the forwarding process.
This complexity can be reduced if an MPLS-like connection-oriented environment is
implemented. In such case packets or bursts are routed over pre-established virtual
network paths.
An important requirement for OBS switch controller is to keep strict the timing
of a burst payload arrival relative to its control packet arrival. Thus the management
of offset times in OBS networks involves additional complexity.
Performance
Several performance parameters can be envisioned. Below a non-conclusive list is
briefly discussed. It contains network utilization, throughput, burst/packet loss probability, and transmission delay.
A disadvantage of OCS is its inefficiency in transportation of traffic which has not
been groomed or statistically multiplexed (e.g., see [CEJ05c][AEBS05][LQYG06]).
Indeed the OCS switching model, which is useful in carrying highly aggregated longlived traffic streams, does not fit well within the Internet paradigm of packet switching. Therefore the network utilization or more particularly link utilization in OBS/
OPS can be higher than in OCS due to the statistical multiplexing that allows for
a better exploitation of network resources and allows to fit to the actual traffic demands. As a result, the amount of network resources (like e.g., number of interfaces
to legacy networks, number of consumed wavelengths) necessary to transport specific
amount of IP traffic might be larger in OCS networks than in OBS/OPS networks.
When considering the throughput of OBS/OPS networks, one has to deal with
the burst contention problem which is further complicated due to the lack of optical
random access memories (RAM). Introduction of fibre delay line (FDL) buffering and
wavelength conversion in OPS networks reduces the packet loss probability. Although
Chapter 2. Optical Switching Architectures
15
similar techniques can be applied in OBS networks, still, considerable burst durations
put limitations on effective employment of FDL buffers in these networks. As a result
we experience high bursts losses in buffer-less OBS networks.
The connection blocking probability is irrelevant for OBS/OPS networks, while it
is an important parameter for OCS networks, operating on circuits. It can be shown
that introducing sparse wavelength converter pools decreases the connection blocking
probability significantly [KA98].
Delays in optical networks may arise due to several reasons. The main transmission delay factor is the propagation delay of optical signal through the optical
medium, i.e., optical fibre. Another delay factor is related to the processing and
switching operations in intermediate nodes. It is envisioned that the delays produced
in intermediate nodes in OPS will be significantly small, to be neglected. In OBS networks, in addition, a burst assembly process may add considerable delays. The delays
caused by the setup of offset times might be either small or long depending on the
switching and processing technologies used in intermediate OBS nodes. The delays in
OCS networks concern mainly delays produced during the connection establishment
process. When the connection is established the propagation delay is dominant delay
factor.
Finally, an important aspect is the performance at the TCP layer. There is no
problem in the case of OCS switching model because the transmitted information does
not experience delay variation or losses in optical domain. On the contrary, both
OBS and OPS architectures need for dedicated TCP implementations to overtake
the degradation of performance due to the higher burst and packet drop rates as
well as a possible problem of burst and packet reordering (e.g., see [SPG05][Gun07]).
Indeed, the ‘edge’ problem of OPS/OBS networks is the out-of-order packet/burst
arrival which is produced either by the contention resolution and QoS algorithms at
intermediate nodes or due to multi-path routing. This issue raises the additional
requirement of supplementary hardware (i.e., large memories) at the destination (i.e.,
sink) nodes due to the reordering operations, which adds further packet transmission
delays.
Flexibility
Flexibility in this context expresses the capability to adapt client signals with different
bit-rates and data formats in order to transport them through the same network
infrastructure.
Appropriate functions are required in edge nodes to adequately adapt different
client signals into a common underlying optical transport network. In the OCS architecture such adaptation is achieved either by using an E/O conversion (if the interface
of the legacy network element is electrical) or with wavelength conversion (using a
transponder/adapter). The adaptation issue in OBS architectures is more complex
and involves the aggregation/assembly function. This function is responsible for burst
formation at the ingress edge node and for the burst disassembly at the egress node
for a given burst flow. The design issue is the proper choice of a burst aggregation
scheme with its associated parameters, namely the size of the burst and acceptable
Chapter 2. Optical Switching Architectures
16
burst formation time.
The granularity in the OCS model is very coarse. It corresponds to the bit rate of
the wavelength and it is determined at the connection time and fixed for the duration
of the connection. OBS/OPS granularity can be very flexible compared to the OCS.
These architectures allow for a packet-level switching, moreover, theoretically each
packet can have a different bit rate.
A scalability factor that expresses the facility of a new connection/path establishment (MPLS paradigm) gives the OPS/OBS architectures an advantage over the
OCS switching model.
QoS
An advantageous feature of connection-oriented circuit switching architectures is that
they have no concept of QoS. In the network there is no need to performing QoSbased queue management, as the necessary and sufficient resources (from ingress to
egress) are assigned prior to the transmission of the actual data. The issue lies in the
fact that there might be contention in the access to transmission resources for the
connection requests of different QoS classes.
In OBS/OPS architectures each individual burst/packet has particular QoS attributes, and thus each individual unit requires to be processed. The QoS attributes
can be encoded in the reservation mechanism or imbedded in each burst/packet as
a code-point that triggers consequent scheduling actions at each intermediate node.
Therefore, additional mechanisms based on properly designed algorithms with the
hardware on node level as well as network QoS mechanism (e.g., QoS differentiation
mechanism, QoS routing algorithm) have to be implemented. These mechanisms
should consider burst/packets prioritization (scheduling), resource reservation, and
admission control capabilities. Therefore, the complexity (and costs) to integrate
QoS in OBS/OPS networks is usually high.
Control plane
GMPLS/ASON has been proposed as control plane (connection management, protection/ restoration, etc.) architecture for OCS networks. Equally, GMPLS/ASON
architecture is considered as the control plane architecture for OBS/OPS networks,
nevertheless some modifications to the proposed architecture are required and necessary to be taking into account to allow the protocols to operate over the OBS or OPS
networks (e.g., see [PSPCK07]).
Signalling overhead (i.e., the volume of signalization/control data) in optical networks is related to the amount of managed data demands. Each circuit switched
connection oriented demand and the OPS packet or OBS burst essentially hold the
same information. In the circuit demand, the source and destination address is signalled, a circuit is established and all data is transmitted over the established circuit.
In the OBS/OPS, each individual burst can be considered as a, although very short,
connection circuit. Thus the highest overhead is observed in OPS networks, where
each optical packet carries header control information (corresponding to the routing
Chapter 2. Optical Switching Architectures
17
tables in each node). Due to the inherent nature of the OBS architectural model the
signalling overhead is significantly reduced. Another aspect is the network’s ability
to adopt to varying traffic demands. In OCS networks, there is significantly more
time necessary to configure the necessary resources when the establishment of a new
light-path is performed, which was let’s say triggered by an increase in traffic demand,
compared to the OBS/OPS approach. In the OPS/OBS network architecture, an increase in traffic demands will require significantly less time for resource configuration,
due to the inherent statistical multiplexing principle.
Under high traffic load network scenarios, however, advanced routing and admission control algorithms are expected to be implemented: firstly, in order to avoid
contention, by distribution of the traffic over the network in a more engineered manner
and thus using the available resources optimally, secondly, to keep the burst blocking
rates in some operational bounds.
For connection-less burst and packet switching, the UNI (user-network), I-NNI,
and E-NNI (network-network) interfaces need to be adapted to pass burst and packet
switching specific information. Besides building the related routing tables, also the
address translation may need to be implemented in case of specific addressing schemes.
For connection-oriented burst and packet switching, the routing of virtual or physical circuits with QoS constraints needs to be supported.
For hybrid solutions (wavelength switching combined with OBS), signalling and
resource reservation concepts need to be elaborated.
Further tasks for the design of the Control Plane are e.g., the distribution of topological information, the used addressing scheme in the OBS/OPS networked layer,
enhanced protections schemes (packet 1+1) for packet based transport technology.
Besides, the design of a suitable DCN (control channel) for the UNI, NNI has to be
defined.
Routing management
In general, the objectives of routing management are the traffic load distribution over
the network (i.e., traffic engineering) as well as preserving the high priority traffic
from the best effort traffic (i.e., QoS routing).
The dynamics of OCS networks are incomparable lower than of OBS/OPS networks. The OCS networks operating in a circuit-switched mode create optical connections for the long-term end-to-end data transmission without wavelength sharing
capabilities. In such networks, we can consider quasi-static network state (i.e., wavelengths/bandwidth occupancy on particular links is known) when we deal with the
so-called routing and wavelength assignment (RWA) problem [ZJM00]. This allows
for solving the routing problem on the basis of actual network state information.
The highly dynamic characteristics of OBS/OPS networks, due to very fast burst
and packet transmission, produces inaccuracy effects in the network state information. This involves network performance degradation (i.e., increased burst/packet
loss probability). Moreover, there is a need to deal with a big number of relatively
small transmission units (bursts/packets) in OBS/OPS networks. It makes the problem more close to the routing in IP networks with the additional issue of the lack
Chapter 2. Optical Switching Architectures
18
of optical memories causing that the switch has to use complex contention resolution algorithms in order to provide an acceptable burst/packet loss rate level when
bursts/packets contention occurs (a related problem is providing of QoS guarantees).
Another issue is the high throughput of optical switching technology that involves
huge requirements for processing capacities of a switch control unit (e.g., lookingup the routing tables). All these factors increase the network complexity an involve
additional functionality requirements. The application of connection-oriented MPLS
environment with its constrained labelled-switched logical paths (for TE and QoS
purposes) and fast labels look-up can support the routing management.
The common objective of routing management in OCS networks is minimizing the
connection blocking probability by means of explicit routing algorithms. Regarding
the RWA problem, the main goal is to establish optical circuits in the network with
optimization of the wavelength resource utilization. Routing algorithms in OBS/OPS
networks have to balance the traffic load (and so burst/packet loss probability) by
means of adaptive routing algorithms in order to help the nodes in the contention
resolution problem.
In the OCS model, some strategies are defined to protect or restore the light-paths
in case of failure while the problem is still not enough addressed in OBS and OPS.
2.3
Summary
According to the specific features of optical circuit, packet and burst architectures, we
can identify their advantageous and drawbacks, which are summarized in Table 2.1.
We can also foresee the application of these switching models in the future network
deployments.
With respect to the flexibility as well as the network utilization and wavelength
consumption, optical circuit-switched networks lag behind compared to optical burst
and packet-switched networks. Nevertheless, technology availability makes OCS
networks an upcoming solution for the next generation transport networks. The
only factor that can delay their introduction by network operators seems to be the
need for regaining the costs of investments of already exploited transmission systems
(SDH/SONET).
The operation of optical packets switching is very effective, particularly, due to
the packet-level transmission and switching granularity. Nevertheless, realization of
such bandwidth-efficient, flexible and data centric all-optical networks faces significant challenges. The complexity of the OPS and the high technological requirements
can significantly shift the development of OPS networks into the future. This prognosis can undergo modifications as some technological breakthrough in the photonics
occurs.
Optical burst switching combines the best characteristics of both OPS and OCS
architectures. The bandwidth granularity of OBS networks lies between the bandwidth granularity of OCS and OPS, and relatively relaxed technological requirements
(especially for switching and processing components) make it an interesting solution
for next generation optical networks. However, the high blocking probability is con-
Chapter 2. Optical Switching Architectures
Advantages
Drawbacks
Foreseen
for the future
19
OCS
OPS
OBS
Natural QoS; reliability
Components & subsystems commercially available
Very
high
flexibility
(traffic dynamics)
Very efficient network
utilization
Reduced node size
High flexibility (traffic
dynamics)
Efficient network utilization
Possible using of lower
speed switching elements
Low flexibility and network utilization
Very high wavelength
consumption, large node
sizes
Only preliminary components & subsystems available
High control complexity
(processing effort. QoS,
routing etc.)
Resilience more complex
Requires more effort for
packet reordering
Components & subsystems partly available
Control complexity (QoS,
routing etc.)
Effort for traffic aggregation
Resilience more complex
A short term deployment
Waiting for technological
breakthrough, especially
for compact, low-cost optical components
A longer term deployment
OBS is a viable solution
for efficient optical networks
A mid term deployment
Table 2.1: Advantages, drawbacks and foreseen for future implementations
sidered a serious challenging issue of OBS. Therefore, there is a strong requirement for
contention resolution mechanisms (wavelength conversion, FDL buffers, and deflection routing, opto-electronic solutions), also with QoS support, applied in hardware
or as an accurately operating control algorithm.
Moreover, routing algorithms should be enhanced with TE functions in order to
alleviate the contention resolution problem by appropriate traffic load distribution
over the network.
Chapter 3
Optical burst switching
The idea of optical burst switching (OBS) has arisen as an alternative to a low-flexible
optical circuit-switching network operation and technological immaturity of optical
packet switching solutions (we have discussed these and some other issues in details
in Chapter 2).
The principal design objective for an OBS network is that aggregated user data
is carried transparently as an optical signal, i.e., without an O/E/O conversion. This
optical signal goes through the switches that have either none or very limited buffering capabilities. Besides, the control information is carried on a dedicated channel
and separately from the user data. In such a network the wavelength resources are
allocated temporally and shared between different connections. It increases network
flexibility and its adaptability to the bursty characteristics of IP traffic. Moreover, the
aggregation of user data helps to reduce the scale of control information processed
in the network as well as it relaxes the switching requirements. Since the control
information and the user data are separated they can be encoded with different modulation formats as well as transmitted at different rates. Such division improves
network management and provides additional flexibility.
Other justification for OBS concept comes from the network user side. Yet not
long ago the predictions on expected services talked mainly about a meaningful participation of real-time multimedia applications with streaming video and broadcasted
TV services in packet networks. Instead, the dominance of multimedia and data
file transfers (e.g., MP3/divx) using various P2P services together with still limited
streaming traffic modifies previous goals [Odl04]. With such P2P services, the typical methods being planned for controlling networks do not fit to user expectations
well. The matter to users now is getting a quite big amount of bits quickly, with low
transaction latency. OBS concept with fast optical transmission of huge amounts of
data seems to match to these expectations well.
Similar objectives of high capacity and usually long-distance data transfers are
in grid networks. A grid network is a distributed collection of heterogeneous computational, storage and network resources (see Figure ??). Most of current operational grids are dedicated to a limited set of computationally and/or data intensive
scientific problems, like e.g., energy physics, astronomy weather forecast or high performance computing/visualization. The requirements of grid applications comprise
21
Chapter 3. Optical Burst Switching
22
among other things high bandwidth transmission, low connection set-up times and
varied transmission granularity for both short and long grid jobs. Network flexibility
and huge optical capacity of OBS technology are the appropriate characteristics for
actual and future grid applications.
3.1
Overview of general OBS concepts
An OBS network consists of a set of electronic edge nodes and optical core, or intermediate, nodes connected by DWDM links (see Figure 2.3). Ingress, or source, edge
nodes aggregate data coming from client networks and assemble them into optical
bursts. Each burst is composed of a data payload and a control packet. The burst
control packet is generated when the assembly process of the burst data payload is
finished. The burst control packet carries all the information necessary to discriminate the burst inside the network, like for instance, the burst class or its length.
In OBS networks there is a strong separation between data and control planes. In
particular, the burst data payload, which is the carrier of user data, is transmitted on
one of data wavelengths, whilst the burst control packet with its signalling message is
transmitted on a dedicated control channel (wavelength). The control channels can
be either out-of-fibre or in-fibre. In the former a dedicated fibre is provided only for
the transmission of control information, whilst in the latter the control channels use
the same fibre as the data channels. Inside an OBS network the control information
is processed electronically, whilst the data burst payload is transmitted all-optically,
without optical to electrical conversion.
In an OBS network the burst control packet is delivered to the core node with
some offset time prior to its data payload. The offset time is introduced in order to
give time for both processing of burst control information and reconfiguration of the
switching matrix. The control packet is processed in an electronic controller of the
switching node. The controller performs several functions, among others the burst
forwarding and resources reservation. The forwarding function, which is related to
the network routing, is responsible for determination of an output link (port) the burst
is destinated to. The resources reservation function makes a booking of a wavelength
in the output link for the incoming burst. In case the wavelength is occupied by
another burst a contention resolution mechanism, if exists, is applied. The contention
resolution mechanism may require a scheduling policy if alternative resources can be
provided for the burst transmission. Also, a quality of service (QoS) provisioning
function, if implemented, may involve particular treatment of higher priority bursts.
In case no resources are available for the incoming burst it is lost.
After the burst transmission is finished in a node the resources can be released for
other connections.
3.1.1
Signalling
OBS signalling adapts the ATM block transfer (ABT) standard proposed for burstswitching ATM networks [I.300]. There are two versions of ABT protocol, namely:
Chapter 3. Optical Burst Switching
core node
ingress node
core node
23
egress node
core node
core node
egress node
burs
burs
offset time
t con
pack trol
et
processing
delay
reservation delay
ingress node
t con
pack trol
et
processing
delay
data
burst
ACK
Time when
resources are
allocated
data
burst
t
t
a) two-way signalling
b) one-way signalling
Figure 3.1: Signalling protocols in OBS networks.
• with delayed transmission, which is known as a tell-and-wait (TAW) signalling
in OBS [Wid95], and
• with immediate transmission, which is called a tell-and-go (TAG) signalling
[Wid95][VS97].
The TAW protocol, which is recognized sometimes as the two-way signalling protocol, performs an end-to-end resources reservation with acknowledgment (see Figure
3.1a) [DKKB00]. In particular when an ingress edge node has a burst ready to be
sent it dispatches a request burst control packet towards the network. If all the core
nodes on the routing path can accommodate the burst the request is accepted and
the ingress node is allowed to go ahead with transmission of the burst payload. Otherwise, the request is refused and the ingress node has either to send another request
later or to drop the burst.
The TAG protocol operates with a one-way signalling and it allocates transmission
resources on-the-fly, a while before the burst payload arrives to a switching node (see
Figure 3.1b) [QY99][Tur99]. In TAG signalling the ingress edge node sends a request
burst control packet and after that, immediately, without receiving any confirmation,
it transmits its data payload. If any core node along the routing path cannot carry
the burst because of its congestion the burst is drooped.
A disadvantage of two-way signalling protocols concerns the latency produced
during the connection establishment process [KB02][WZSZ03]. For this reason the
TAW signalling is oriented more towards metro networks. In such networks, short
transmission distances allow keeping low the connection setup times.
Chapter 3. Optical Burst Switching
24
The one-way reservation signalling model allows operating in large-distance networks. In such architectures the problem of synchronization between the burst control
packet and its data payload arises in the network. Therefore each switching node has
to keep updated the information about relative time-scale position of the control
packet and the payload. Another issue is the problem of burst contention in the
network. Indeed, a burst is released towards the network even it is not guaranteed
there are transmission resources available to deliver it to the destination node. For
this reason several contention resolution mechanisms have been proposed to alleviate
this problem, as we discuss later.
A great feature of OBS concept is the possibility of operating with two-way and
one-way signalling protocols simultaneously. In particular, in a two-way resources
reservation mode one can setup aside some wavelengths to be used as in an OCS
scenario, whilst one-way reservation messages make a statistical use of the rest of
available resources. In this way the same optical infrastructure can support both
static (by wavelength switching) and dynamic (by burst switching) traffic.
In order to make a distinction, our further discussion concerning OBS assumes a
one-way signalling protocol, whilst a two-way signalling model is reflected well in an
OCS network.
3.1.2
Architectures and functions of OBS nodes
An OBS edge node
An inter-working function between client networks and an optical OBS network is
provided in OBS edge nodes (see Figure 3.2). The client network can be any legacy
network like e.g., IP, ATM, SONET/SDH, or other network. An ingress edge node
is responsible for adaptation of the client data signals to the format used in the OBS
network; accordingly, an egress edge node performs the opposite operation.
A few functions can be distinguished to be performed by an OBS ingress edge
node:
• aggregation of data from client networks,
• assembly of a burst payload,
• generation of a burst control packet,
• (optionally) set-up of an offset time,
• burst transmission,
• other functions (e.g., burst segmentation).
Data from client networks is aggregated according to a forwarding equivalence
class (FEC). Each FEC describes client data of similar or identical characteristics,
like for instance their destination, QoS class, or transmission time window. A burst
payload is assembled from the data of the same FEC and according to a given burst
Chapter 3. Optical Burst Switching
25
Offset management
Control packet
generator
Burst
scheduler
Burst assembrer
Timers
Switch
Class 1
...
Classifier
Length
thresholds
Output
links
Destination #1
...
Input
traffic
sources
...
Class K
Burst assembler
Destination #D
Figure 3.2: OBS ingress edge node.
assembly algorithm. The algorithm takes a decision about when the burst aggregation
process should be finished. Several burst assembly algorithms have been proposed for
OBS networks (e.g., see [RG04][YLC+ 04]):
• timer-based - define the maximum amount of time for the burst assembly
process,
• burst length threshold-based - specify the maximum, permissible length of the
burst,
• hybrid timer/length-based algorithms,
• other algorithms (e.g., with exponentially distributed burst inter-arrivals guaranteed).
A burst assembly algorithm influences the overall network performance. In fact
it allows a network designer to control the burst traffic characteristics, e.g., such
as burst arrival process to core nodes or burst length distribution. Timer-based
algorithms have the burst inter-arrival times determined, whilst length thresholdbased algorithms have the burst lengths determined.
An edge node should equip the burst control packet in the burst relevant information, sufficient to handle the burst payload in core nodes. An exemplary burst
control packet, shown in Figure 3.3, comprises information about the burst duration,
the payload arrival time (relative to the control packet arrival), the class of burst and
Chapter 3. Optical Burst Switching
QoS
Message type
Burst arrival time
26
Input identifiers (port, wavelength, label)
Burst duration
other functions
FEC/CRC
Figure 3.3: Burst control packet format.
some routing/forwarding information (input wavelength, an identifier of the routing
path).
In a common OBS scenario the ingress edge node introduces an offset time between
the burst control packet and its payload. In the simplest case such offset is fixed
and equal to the time necessary for processing and switching operations in all the
nodes laying on the longest routing path. The problem of offset time provisioning is
addressed later in more details.
A burst segmentation is another (optional) function that can be found in the
edge node. This function performs a partition of the burst payload onto several
data segments. In case the burst collides with another one in a core node its data
contending segments can be dropped.
An OBS core node
A transparent switching/routing of optical bursts from one fibre link to another is
performed in an OBS core node.
The following functional blocks of an OBS core node can be distinguished (see
Figure 3.4) [[Nor03]]:
• an input interface,
• an electronic switch controller,
• an optical switching core, and
• an output interface.
Main function of the input interface is the extraction of control and data channels.
Each control channel is connected to a burst mode receiver. The burst mode receiver
retrieves the control information from control packets, converts it to electrical form
and passes it down to the switch controller. Simultaneously, the data bursts carried on
different wavelengths are de-multiplexed and delivered to the optical switching core.
Some OBS architectures possess a fibre coil element introduced into the data path.
The fibre coil provides some offset time for the control packet processing operation.
The input interface also monitors incoming signals and conditions them as required,
e.g., through power equalization and regeneration.
The switch controller processes control packets. In particular, it makes a forwarding table lookup, and reserves transmission resources for the incoming data payload.
Chapter 3. Optical Burst Switching
27
Electronic switch controller
Control packet
processors
Output
buffers
...
Input
links
Input
buffers
...
Control
channels
Output
links
Switch fabric
Output
interface
Wavelength
converters,
FDL buffers
pool
...
...
Input
interface
Figure 3.4: OBS core switching node.
The resources reservation is preceded by identifying a suitable switching matrix path
and resolving the contention problem, possibly with some QoS policy introduced. The
controller usually updates information encoded in the control packet. It is responsible
also for sending, in proper instants of time, control signals to the switching core and
other switch components in order to handle optical data bursts.
The optical switching core is built with a switching matrix and other dedicated
components. The switching matrix can be characterized by the mode of its operation
(asynchronous/synchronous), dimension, switching time, internal blocking properties
(e.g., non-blocking) and signal degradation (e.g., see [PPP06]). The dimension of
switching matrix should be (N xW )(N xW ) if N is the number of output/input ports
and W is the number of wavelengths per port (link). Other components that can
be found in the optical switching core are e.g., wavelength converters and fibre delay
lines (FDL); they are used as burst contention resolution mechanisms.
The output interface implements an update of control information, DWDM multiplexing of data and control channels and conditions for optical output signal.
3.1.3
Offset time provisioning
An important feature of OBS architectures is provisioning of offset times, which
separate the burst control packets and their payloads. The offset time gives some
delay budget for processing and switching operations in core nodes, without the need
for buffering of optical data burst payload. The burst is lost if effective processing
Chapter 3. Optical Burst Switching
28
time, the control packet undergoes in the controller, is lower than the delay budget.
Therefore appropriate setup of offset times is crucial in OBS networks.
The offset time can be introduced, either
• in an electronic ingress edge node, by delaying the transmission of data burst
payload ([QY99]), or
• in an optical core switching node, by means of an additional fixed-length fibre
delay element introduced into the data path (as e.g., in [AST+ 06]).
Three different offset-time provisioning architectures can be distinguished, with
regard to the place where the offset time is introduced (see Figure 3.5):
• a conventional OBS (C-OBS), with processing offsets introduced in edge nodes,
• an offset time-emulated OBS (E-OBS), with processing offsets introduced in
core nodes,
• a hybrid OBS (H-OBS), with processing offsets introduced both in edge and
core nodes.
Later we can distinguish four models of offset-time provisioning in OBS networks
with respect to the changes of delay budget a burst experiences during its trip through
the network:
• the delay budget decreases - proper to a C-OBS architecture,
• the delay budget is fixed - proper to an E-OBS architecture in which the burst
control packet is released together with its data payload in consecutive core
nodes (OPS-like operation),
• the delay budget increases - proper to an E-OBS architecture in which the burst
control packet is released immediately after its processing in each core node,
• the delay budget fluctuates - proper to a hybrid architecture.
In C-OBS architectures the offset time is setup in a soft-way, by delaying the
transmission of data burst payload with respect to its control packet. The offset
should compensate all switching and processing times of all the nodes lying on a
routing path; hance it can be seen as a global offset, which is setup only once. An
important property of C-OBS architectures is that the offset varies inside the network.
Indeed it decreases after each hop by the time the control packet spends in the node
controller.
In E-OBS architectures the offset time is introduced in a hard-way, by means of
a fibre delay coil element which postpones the arrival of the data burst payload to
the switching matrix. The fibre delay coil is a passive piece of fibre of fixed length.
Such element is responsible only for compensating the switching and processing times
produced in the corresponding node; the offset is local and it has to be introduced in
each switching node. On the contrary to C-OBS, in E-OBS it is possible, in principle,
to keep the offset times fixed in consecutive nodes.
Chapter 3. Optical Burst Switching
29
a) C-OBS
ingress node
control
packet
burst payload
d
core node
D
core node
D
egress node
D
t
b) E-OBS
ingress node
control
packet
burst payload
core node
d
D
core node
d
D
egress node
d
D
t
c) H-OBS
ingress node
control
packet
d
burst payload
core node
d
D
core node
D
egress node
D
t
d - introduced offset time
D - processing delay
Figure 3.5: Offset time provisioning architectures.
Chapter 3. Optical Burst Switching
3.1.4
30
Resources reservation
A resources reservation process in the core node concerns the reservation of resources
necessary for undisturbed switching and transmission of data bursts from input to
output ports. This process includes reservation of switching resources (i.e., a switching
path in the switching matrix), a wavelength in the output link, and other shared
resources, e.g., wavelength converters or FDL buffers, depending on capabilities the
node is enhanced with.
Separation of data and control channels together with the offset-time provisioning
allows using different resources reservation schemes in OBS networks. Each reservation starts from the setup and finishes after the resource release entity. Both resources
setup and release can be either explicit or estimated [BP03]:
• explicit setup - the resources are configured immediately upon processing of the
control packet,
• estimated setup - the reservation of resources and configuration of the switching
matrix is delayed until the actual burst arrival,
• explicit release - the source sends an explicit trailing control packet to signify
the end of a burst transmission,
• estimated release - the end of a burst transmission is known from the burst
length, and therefore the moment of release can be calculated.
Different resources reservation algorithms have been proposed by adopting the
above presented rules:
• Just-In-Time (JIT) [WM00] - performs an immediate resource reservation (see
Figure 3.6a). It checks for the wavelength availability just at the moment of
processing of control packet. It adopts either explicit or estimated resources
release. The advantage of this algorithm is its simplicity, however, at the cost
of worsen efficiency due to the overprovisioning of resources.
• Horizon [Tur99] - performs estimated setup and resources release. It is based
on the knowledge of the latest time at which the wavelengths are currently
scheduled to be in use.
• Just-Enough-Time (JET) [YQ97] - performs estimated setup and resources release (see Figure 3.6b). It reserves resources just only for the time of burst
transmission. It is one of the most efficient mechanisms, with improved burst
blocking probability when comparing to other algorithms. A disadvantage of
JET algorithm is its high complexity as long as it allows for filling the voids
that occur between already performed reservations.
In case of an estimated resource reservation the control packet should carry exact
information about burst payload arrival and its length.
Chapter 3. Optical Burst Switching
31
a) Just-In-Time
ingress node
core node
core node
offset
burst
setup
Processing
delay
release
Switch
configured
setup
release
egress node
t
b) Just-Enough-Time
ingress node
offset
burst
setup
core node
core node
Processing
delay
setup
Switch
configured
setup
delay
release
release
egress node
t
Figure 3.6: Resources reservation schemes.
3.1.5
Contention resolution
A burst contention occurs when more then one burst solicit for the same resources
at the same moment. Resolution of the burst contention is a crucial problem in OBS
networks. Two factors that complicate the contention resolution are unpredictable
and low-regular burst statistics [LES00] and the lack of optical memories. Loosing
a burst that aggregates a number of packets may have worse effect than loosing a
single packet. The case might be really serious if the burst carries packets belonging
to TCP connections [CR06].
Similarly like in OPS networks, the burst contention can be resolved with the
assistance of following mechanisms (see Figure 3.7):
• wavelength conversion (WC) [ELP03] - converts the frequency of a contending
burst all-optically to other, available wavelength;
Chapter 3. Optical Burst Switching
a) Wavelength conversion
32
b) Fiber Delay Line Buffering
burst 1
Incoming bursts
Outgoing bursts
l1
burst 2
burst 1
l1
burst 2
l2
l1
l2
Fiber Delay Line
Wavelength
converter
c) Deflection routing
d) Burst Segmentation
Primary paths
burst 1
burst 2
Incoming
bursts
Contention
Deflective path
burst 1
burst 2
Dropped
segments
Figure 3.7: Contention resolution mechanisms.
• deflection routing (DR) [CZZ04] - forwards a burst spatially, in the switching
matrix, to another output port (fibre).
• fibre delay line (FDL) buffering [HCA98] - operates in time domain and resolves
the contention by delaying the departure of one of bursts by a specific period
of time.
In case none of mechanisms can resolve the contention the burst is dropped.
The wavelength conversion is a natural way to resolve contentions in OBS networks. A drawback of this mechanism, however, is high cost of WC devices, especially,
in case of a full-wavelength conversion, which is performed in the wide frequency
range.
Application of the deflection routing in OBS networks is almost cost-less since
no additional devices are necessary in order to run this mechanisms. On the other
hand, the operational complexity may be high since the mechanism should assure
that a deflected burst reaches its destination, even when forwarded to another node
output link. Efficiency of this mechanism depends heavily on the network topology
and routing strategy as long as the contention is resolved by re-routing the traffic
to adjacent nodes. It was shown that the deflection routing can improve network
performance under low and moderate traffic loads whilst it may intensify burst losses
under high loads [ZVR+ 04]. Another difficulty that has to be managed properly
is the out-of-order burst arrival problem, which occurs when bursts traveling over
variable-length paths arrive to the destination in a disordered sequence.
Chapter 3. Optical Burst Switching
33
Even if one of the principal design objectives for OBS was to build a buffer-less
network the application of FDL buffering is considered as well. Both feed-forward
and feed-back FDL buffer architectures can be used [Gau02]. In [Gau03] it was shown
that combined application of FDL buffering with WC can significantly reduce burst
losses in OBS networks. As a buffering tool the FDLs are bulky and not scalable.
Comparing to electronic buffers and their role in current packet networks, the FDL
offers only a limited buffering capability. For a typical fibre span of 80km length
the corresponding maximum delay that can be introduced by FDL buffer, without
need for optical signal amplification, is 266µs [Gau03]. In order to work effectively an
FDL buffer has to provide several delays. Therefore a basic delay unit of such buffer,
which for efficiency reasons should correspond to the average burst duration [CC01],
can not exceed some tens of µs. Considerable burst durations significantly limit the
application of FDL buffers in OBS networks, when comparing to OPS networks.
Another technique that aims at the reduction of data losses is a burst segmentation
technique [VJ02a]. In this mechanism each burst is divided into a few segments.
If a burst contention occurs, instead of loosing the entire burst either the head or
tail segments of one of the contending bursts are dropped. The burst segmentation
increases the operational complexity due to additional information about the burst
segments that have to be transported and processed. Moreover, when the node loses
some packets of the burst head and the following packets in the same burst belong
to the same flow, it breaks the correct packet sequence what causes a degradation
problem at the end-to-end transport protocol (such as TCP).
3.1.6
Burst scheduling
A scheduling algorithm undertakes a decision which wavelength or FDL delay has
to be assigned to a given burst in case there are more resources available. The simplest scheduling schemes can be based on either a random or a round-robin resources
selection.
More advanced scheduling policies, which are based on Horizon and JET resources
reservation mechanisms (see Figure 3.8), are:
• the latest available unused channel (LAUC) [XVC99], which is a Horizon-type
algorithm; it keeps a track of the latest unscheduled resources and searches for
a wavelength with the earliest available allocation;
• the void-filling (VF) [XVC00], which is a JET-based algorithm; it keeps a track
of the latest unused resources and allows putting short bursts into time gaps
before the arrivals of future scheduled bursts. Several variations of the VF
algorithm can be found in literature (e.g., see [MRZ04]). A VF algorithm can
achieve better performance than a Horizon-based one, however, at the cost of
high processing complexity. In an OBS with FDL buffering, this complexity
can be decreased when using a FDL-batch algorithm [XQLX03].
There is a group of scheduling techniques which are enhanced with so called lookahead processing window capability (e.g., see [FJ03a][JECA03]). The look-ahead
Chapter 3. Optical Burst Switching
34
new burst
previous reservations
void
burst
dropped
LAUC
void-filling
Figure 3.8: Burst scheduling algorithms.
window allows collecting more information about incoming burst reservations, and
thus their optimized processing can be performed.
3.1.7
Quality of service provisioning
OBS architectures need for dedicated QoS mechanisms in order to preserve the
quality-demanding applications from the best-effort data traffic. Since optical networks do not have an equivalent to electronic random access memories the problem of
burst-loss quality guarantees is very challenging. On the other hand, almost buffer-less
and fast transmission in OBS networks may result in lower latency than in traditional
data networks.
The detail of QoS provisioning in OBS networks as well as some particular issues
are addressed in Part 3 of the thesis.
3.1.8
Network routing
OBS architectures with no buffering capabilities are sensitive to burst congestion. A
proper routing strategy, enhanced with some traffic engineering (TE) capabilities,
may help in the congestion reduction. A highly dynamic character of burst traffic, however, may result in the inaccuracy of network state information. Moreover,
there is a need to deal with a big number of relatively small data bursts. Other issue is the high throughput of optical switching technology which involves additional
requirements for processing capacities of switch controllers (e.g., fast looking-up of
routing tables). All these factors increase network complexity and involves additional
funcional requirements in OBS.
Application of a connection-oriented multi-protocol label switching (MPLS) architecture [Ros01], with its explicit routing and fast labels look-up, can help in the
discussed issues. As a consequence several routing methods (e.g., see [ZLW+ 04]
[ZWZ+ 04][LY06][HHM05]) apply the concept of labelled OBS (LOBS) for TE, as
Chapter 3. Optical Burst Switching
35
proposed in [Qia00].
An important issue related to the routing problem is end-to-end QoS provisioning; in this context, several solutions have been proposed in the literature (e.g., see
[VJ02b][LKSG03][KG03a][LYH+ 06][ACP04]). Some routing strategies support network resilience by the computation of backup paths ([CMC06][Bou03][HHM05][GZ06]
[JQX00]). A study on multicast routing in OBS networks can be found in [JXC+ 00]
and [JQX00].
Part 4 of the thesis is devoted to the routing problem in OBS networks.
PART II
Offset Time-Emulated OBS
Architecture
Chapter 4
E-OBS architecture
From the very beginning there have been considered two distinct concepts of offset
time provisioning in OBS networks [QY99]. In a conventional OBS (C-OBS) the offset
time is introduced in the edge node by delaying the transmission of the burst payload
with respect to its control packet. On the contrary, in an offset time-emulated OBS
(E-OBS) the offset time is provided in each core node by means of additional fibre
delay element. Although C-OBS has attracted lots of attention it possesses many
disadvantages that can be avoided in E-OBS.
The intention of this Chapter is to point out the strengths and weaknesses of COBS and E-OBS architectures. At the beginning we introduce operational principles
of an E-OBS architecture. Then we lead comparative discussion on several issues
related to both functional and performance characteristics of E-OBS and C-OBS.
4.1
4.1.1
Principles of E-OBS
Node architecture
E-OBS core node is a typical OBS node (e.g., [XVC00]) with additional pool of fibre
delay coils (FDC) introduced into the data path of the input interface (see Figure
4.1). The control channels are provided either out-of-fibre (i.e., in a dedicated fibre)
or in-fibre (i.e., in the same fibre as data channels). In the case the in-fibre control
channels are used, they should be filtered before the pool of FDCs. To perform this
function a passive optical device like e.g., a band splitter module, or an optical drop
multiplexer (ODM) can be used.
The input control channels pass through the optical to electrical (o/e) converters
and are directed to the switch controller. The controller is equipped with input buffers
to store the incoming control packets before their processing in the processor unit(s)
(CPU). After that and some output buffering, control packets are converted back to
the optical signal form and transmitted through the output control channels to the
output interface. If in-fibre control channels are used, the output interface combine
both data and control channels in a multiplexer into an output fibre.
The input data fibres, after separation of control channels, pass through the pool
39
Chapter 4. E-OBS architecture
40
routing
table
Out-of-fiber Control
Channels (CC)
Controller
input CPU
buffers
o/e
output CC
buffers
CPU
e/o
In-fiber
CCs
Filter
…
DMUX
…
Input
fibers
Pool of
Fiber
Delay
Coils
Optical Switch
Fabric
MUX
…
Output
fibers
Filter
MUX
…
DMUX
Figure 4.1: General E-OBS core node architecture.
a) A Fibre Delay Coil
b) A pool of FDCs in an OBS test bed
Figure 4.2: Fiber delay coils; a) a single component, b) a part of an OBS test-bed.
(courtesy of Newport Corp., and OITDA)
of FDCs - each data fibre passes through one FDC. Then the data channels (wavelengths) are de-multiplexed and the data bursts, from each data channel, undergo
all-optical switching operation in the switch fabric to appropriate output ports.
The complexity of FDC is much lower than of any FDL buffer. In fact the FDC
is a piece of fibre of quite limited, fixed length and it does not require any switching
capability. Such components are commercially available (e.g., see [Fib07a][Fib07b]);
exemplary parameters of a FDC presented in Figure 4.2a are: the insertion loss <
0.3db/km, fibre length up to 4km what gives 20µs of delay, operating wavelengths
1260 ∼ 1650nm, dimension 6.0000 × 6.0000 × 1.5900 with enclosure.
It worths to mention that there is a need for only one FDC per each input port.
The maximum nodal degree in the most referenced mesh network topologies (see Section 10.1 for more details, and also [RFL05]) does not exceeds 5, thus the introduction
of a pool of FDCs into an OBS node should not cause much troubles. Indeed some
OBS test-beds operating with FDCs (see Figure 4.2b) are already available [AST+ 06].
Chapter 4. E-OBS architecture
41
data burst
Input Data
Channel
TS
burst control packet (BCP)
Input Control
Channel
lb
Lb
queuing delay processing
LBCP
WCPU
TP
WCPU
*
TFD
JIT reservation
Reserved
Wavelength
Horizon reservation
Output Data
Channel
idle BCP waiting time
Output Control
Channel
time
Figure 4.3: Time dependencies in E-OBS.
4.1.2
Control operation
In principle, E-OBS is considered to operate with a one-way signalling. In a proposed
E-OBS scenario we assume the burst control packet is dispatched from the edge node
prior to its data payload with a small offset introduced just to compensate the switch
re-configuration delay.
When the burst reaches a core node, the control packet goes directly to the switch
controller, whilst the payload is delayed in FDC by some fixed time. During this
period the control packet is queued and processed in the controller so that to reserve
the switching and transmission resources for the arriving data payload. This operation
is repeated in each core node. When the burst reaches its egress node it is disassembled
and data are delivered to the higher layer protocol.
The control packet after its processing can be forwarded to the next node either
immediately or it remains in the controller memory until the processing offset expires
(an example is presented in Figure 4.3). Both solutions have their advantages and
disadvantages.
• In an immediate control packet forwarding, the offset time between the
control packet and the payload increases hop-by-hop what may result in two
effects. On the one hand the increasing offset gives more chances to reserve the
resources and it may help the burst to accomplish its trip (e.g., see [YQD01]
[KCM04]). On the other hand there might be additional waste of resources in
the case of JIT resources reservation since the reservation periods increase in
consecutive nodes.
• In a delayed control packet forwarding, the time distance between the
control packet and the payload, in principle, is kept fixed from link to link
Chapter 4. E-OBS architecture
42
control packets
Flow1
- fail
Flow2
- fail
Flow3
- success
D4
CN1
Flow 1
Flow 2
bursts
CN2
D3
D2
Dx: processing time
at node x
CN3
CN4
Flow 3
Figure 4.4: Unfairness in conventional OBS.
inside the network. Thanks to this feature there is no variation of offset times
in E-OBS. The only inconvenience is a possible contention of control packets in
the output control channel. Therefore some output buffers has to be used and
the emerging buffering delays should be compensated by the FDC.
In this thesis we concern on the delayed control packet forwarding.
An important requirement for the controller is to keep strict the timing of a burst
arrival relative to its control packet arrival. We assume that on entry to the switch
each burst control packet is time-stamped. Then after its processing and scheduling
to an output queue the relative burst arrival time is re-calculated and updated in the
control packet.
4.2
Characteristics of E-OBS and C-OBS
Fairness
In conventional OBS, whilst the control packet is forwarded through the network its
global offset time decreases successively at each hop by processing time, which is the
time the control packet spends in the node controller. The emerging variation of
offset times can produce unfairness in access to transmission resources (see Figure
4.4). Indeed a burst of higher number of hops remaining to reach the destination,
and thus of larger offset time, has more chances to reserve an output wavelength than
a burst of smaller offset time. The described effect starts to play role if the offsets
assigned to the bursts are larger than the burst durations (e.g., see [DG01]). It worths
to mention that this feature has been used in an offset-time differentiation mechanism
designed for QoS provisioning [YQD01].
Chapter 4. E-OBS architecture
43
1,E+00
Burst Loss Probability
1,E-01
EON
1,E-02
1,E-03
NSFNET
1,E-04
1,E-05
C-OBS
E-OBS
1,E-06
1
2
3
4
5
6
7
8
remaining hops # to the destination
Figure 4.5: Burst loss probability vs. remaining hops number.
Another negative aspect related to the unfairness is the path length priority effect
(see [KCM04]). This effect corresponds to the increased loss probability of bursts
that approach their destination and at the and of their trip have small offsets. In
particular such bursts can be easily overtaken by the bursts of higher offsets, e.g.,
which have just been expedited from the ingress node. As a consequence, we could
have unnecessary waste of transmission resources that were already utilized in all the
nodes traversed by the lost bursts.
In order to illustrate the unfairness effect, in Figure 4.5 we present some exemplary
simulation results. We consider two network topologies called NSFNET (an American
backbone network) and EON (a pan-European network) of 15 and 28 nodes, and
23, and 39 links respectively (see Section 10.1 for more details about the network
scenarios). Each link has 32 data wavelengths and the transmission rate is 10Gbps.
Each node is an edge node generating 25.6Erlangs (0.8 load, when normalized to the
link capacity). Bursts have exponentially distributed inter-arrival times and lengths
(mean duration of 32µs). 1µs and 10µs are the times considered for the switching and
processing operation respectively. The JET resources reservation with the LAUC-VF
scheduling is used. Shortest path routing is applied.
We can see that the bursts that begin their trip, i.e., of high number of remaining
hops to the destination, undergo lower losses than the bursts which have just the
ultimate hops to reach the destination in C-OBS. On the other hand, in E-OBS each
burst has the same chances to reserve the transmission resources as long as the offset
times that are determined by the length of a FDC are the same. The results presented
in 4.5 confirm this observation. In particular, we can see that the burst loss probability
Chapter 4. E-OBS architecture
44
1,E+00
Overall Burst Loss Probability
C-OBS
E-OBS
1,E-01
EON
1,E-02
NSFNET
1,E-03
1,E-04
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
Offered load
Figure 4.6: Burst loss probability vs. offered traffic load.
is much more stabilized without regard to the number of hops remaining to reach the
destination; its slight variation is due to an unbalanced traffic-load distribution. As
a consequence, the unfairness in access to transmission resources experienced by the
bursts belonging to different connections disappear in E-OBS.
Burst loss and delay performance
Figure 4.6 compares the overall burst loss probability obtained as a function of offered
traffic load (normalized to the link capacity) under different network scenarios. As
we can see both C-OBS and E-OBS architectures offer similar performance.
Transmission delay produced in OBS networks is due to the link propagation
delay dl (approx. 1ms in 200km link) and the offset provided for the processing δp
(up to several µs) and switching δs (below µs if a fast switching matrix is used, as
e.g., in [GWL+ 05][AST+ 06]) purposes. We have already assumed that in E-OBS the
switching offset between the burst control packet and the data payload is introduced
in the edge node. Hence, the end-to-end transmission delay D the burst undergos is
the same in both C-OBS and E-OBS architectures, and it can be expressed as:
D = (n + 1)dl + nδp + δs
where n is the number of intermediate switching nodes
Note that the propagation time is still a dominant delay factor.
(4.1)
Chapter 4. E-OBS architecture
45
a) JIT in C-OBS
Source
b) JIT in E-OBS
Destination
Source
Control
packet
Control
packet
Burst
payload
Burst
payload
Destination
Figure 4.7: JIT resources reservation in a) C-OBS, and b) E-OBS.
Resources reservation
The Just-in-Time resources reservation algorithm performs an immediate resource
reservation as it checks the wavelength availability just at the moment of the control
packet processing. On the contrary both the Horizon and the JET perform a delayed
resources reservation for the period of time starting from the burst arrival time. The
difference between these algorithms is that the Horizon searches for a wavelength that
does not have any later reservations while the JET allows for filling the voids that
occur between reservations.
E-OBS can operate with any resources reservation algorithm. Nevertheless, the
JIT and the Horizon algorithms seem to be the most appropriate ones, whilst the
JET algorithm is frequently considered for C-OBS networks.
The advantage of JIT is its low complexity since the only information that has to
be kept record of in network nodes is whether a wavelength is currently available or
not. Over-provisioning of resources due to early reservations is the main drawback
of JIT (see Figure 4.3). As a result, burst losses can occur even when there are no
transmission conflicts on the same wavelength. Nevertheless, the reservation periods
are shorter in E-OBS than in C-OBS due to smaller offsets times (see Figure 4.7).
Hence we can expect that the over-provisioning effect of JIT algorithm will have much
lower impact on the performance in E-OBS than in C-OBS.
E-OBS, in principle, does not experience the offset variation inside the network.
If we consider the switching nodes are not enhanced with FDL buffering, a resources
reservation algorithm does not need to be void-filling aware. For this reason the
Horizon algorithm can be used instead of the more complex JET algorithm without
any degradation of performance.
Chapter 4. E-OBS architecture
46
Burst scheduling
One of the key problems of OBS is to schedule the bursts efficiently so that the
throughput is maximized and the burst losses are minimized. In the OBS networks
without FDL buffering the performance of an online best-effort scheduling algorithm
depends among other things on the offset time and the burst length distributions.
In particular, the best performance is achieved when all bursts have the same offsets
and the same lengths [LQXX04]. Whilst C-OBS is characterized by variable offsets,
E-OBS can provide fixed offset times.
Another benefit from the core node-introduced offsets in E-OBS is some facility in
the application of look-ahead processing window techniques. These techniques need
for soma extra offset in order to constitute a look-ahead processing window. The
processing window allows for more efficient burst scheduling in both contention resolution [FJ03a][JECA03] and QoS provisioning [FJ03b][KCMSP06]. Since the processing window can be provided easily in E-OBS, by means of additional FDC delay, its
introduction in C-OBS may seriously aggravate the unfairness.
Quality of Service provisioning
Several strategies have been considered in the literature to provide QoS capabilities
in OBS networks. Among them a burst preemption technique (e.g., see [KA03]) and
an offset-time differentiation technique [YQD01] can offer the utmost performance
with regard to the class differentiation (see Chapter 7 for a detailed analysis). The
former allows overwriting the resources reserved for LP bursts by HP reservations in
case of burst conflicts. The latter assigns an extra offset time to HP bursts, what
favors them during the resources reservation process.
The general drawback of burst preemptive-based mechanisms in OBS is the overbooking of resources in the downstream nodes in case of a successful preemption.
Therefore there is a need for additional signalling procedure to be used in order to
release them, or the resources are wasted. This problem is addressed in Chapter 8,
and we show that the overbooking of resources can be effectively avoided in E-OBS
nodes enhanced with the processing window capability.
Performance of the offset-time differentiation mechanism may be affected by the
multiplication of effective classes due to the offset variation [DG01]. In order to
diminish this effect the offset times should be low enough in C-OBS. E-OBS does not
have such limitations thanks to its fixed offset-time provisioning.
Routing and network survivability
C-OBS architectures have some difficulties with alternative/deflection routing. In
particular, edge nodes should know the routing path prior to the control packet transmission in order to calculate and setup the offset times accurately. When allowing
for alternative routing inside the network, an insufficient offset problem may emerge.
Indeed in case an alternate route is longer than a primary route the burst is dropped
if the control packet does not have enough time to reserve resources ahead of the data
burst. For this reason, the offset time should be either calculated for the worst case,
Chapter 4. E-OBS architecture
47
i.e. for the longest possible alternative path, what may result in superfluous burst
delay, or additional hardware (an output FDL like in [HLH02b]), or control [CEJ05a]
mechanisms have to be involved in order to diminish this effect.
Some OBS restoration schemes presented in the literature consider deflection routing to coop with link failures (e.g., [XTGE+ 04], [HSE04]). Again, an important factor
that has to be taken into consideration here is the insufficient offset effect. Therefore,
the choice of the offset time is very critical due to its influence on the burst losses in
OBS networks.
In E-OBS the routing path can be created freely inside the network with any
alternative routing algorithm as long as the offset time is introduced in each core
node by means of the inlet FDC.
Hardware complexity
Fibre delay coil There is some additional hardware complexity in E-OBS due to
the need for FDCs to be introduced at the input ports of core nodes (we have already
discussed this in Section 4.1). Typical FDC delays necessary for E-OBS operation
range from some µs to tens of µs, depending on switching and control processing
technologies (e.g., see [BBE+ 05]) used as well as particular choices for control algorithms (resources reservation, scheduling, etc.). Therefore considered lengths of FDC
can be between 1 ÷ 5km, as e.g., in [AST+ 06].
The attenuation of optical signal (below 0.3dB/km) should be taken into account
when analyzing the power budget and designing the amplification stages. It is important to say that there is a need for only one FDC per node input port which
compensates offset-times for all the data channels simultaneously. The control channel should be extracted before the FDC module and brought to the switch controller.
The application of FDC might be advantageous in the context of signal regeneration
since this fibre could act as a dispersion compensation unit for the optical signal
entering the node.
Memory requirements The requirements related to the amount of electronic
memory installed in C-OBS nodes are higher than in E-OBS nodes. C-OBS edge
nodes require for output electronic buffers to store the assembled bursts for the offset
period. The capacities of such buffers greatly depend on the burst assembly parameters as well as on the offset times itself. In some OBS scenarios the burst payloads are
considered to carry some M bytes of data. Moreover the offsets, which comprise the
processing times for all core nodes laying on the routing path, might be very large.
As a result the memory requirements in C-OBS might be really high.
In E-OBS the burst after its assembly has to wait in the edge node only for a short
switching offset period. Then it is sent towards the network as soon as there are free
transmission resources in the output link. Some additional buffers are necessary in EOBS switching nodes, because the burst control packets might need to be stored after
their processing. Nevertheless, the memory requirements in this case are moderate
as long as the lengths of control packets are very small.
Chapter 4. E-OBS architecture
Fairness
Performance
Resources reservation,
scheduling complexity
QoS
Alternative routing
Hardware complexity
48
C-OBS
E-OBS
No
Yes
BLP slightly better in E-OBS, end-to-end delay the same
High
Low/Medium
Some difficulties
Limited
Memory (at the edge)
Some facilities
Not limited
Fibre delay element (in the core)
Table 4.1: Advantages and drawbacks of offset-time provisioning architectures
4.3
Summary
Table 4.1 summarizes both the qualities and drawbacks of the discussed offset-time
provisioning architectures. The E-OBS surpasses the C-OBS in many aspects as
we discussed it in this Chapter. Therefore there is a motivation for recognizing the
E-OBS as an efficient and functional solution for OBS networks.
Although not mentioned before, there is one more great benefit of E-OBS. Thanks
to the application of fiber delay elements, which makes the operation of E-OBS and
OPS are very similar, an E-OBS architecture can be seen as an immediate migration
step towards OPS architectures. Still, the differences between both technologies lay in
the length of transmission units (burst vs. packet) and signalling mode (out-of-band
vs. in-band), and thus, higher hardware and processing requirements of the OPS.
Nevertheless, as the progress in optical technologies will continue these dissimilarities
should disappear in the future. As a result, the application of E-OBS may facilitate
the migration from OBS networks to OPS networks.
Since E-OBS architectures need for additional fibre delay elements, the study on
their feasibility is provided in the next Chapter.
Chapter 5
Modelling of E-OBS control plane
5.1
Introduction
Due to the separated transmission of burst control packets and data payloads both
opto-electronic control and all-optical data planes can be seen as two parallel networks, namely a data and a control network (see Figure 5.1). The burst is lost if
either its control packet or its payload is lost; it occurs, in principle, due to resources
occupancy in congestion states. Both burst contention resolution mechanisms and
scheduling algorithms deal with the problem of congestion in data plane (e.g., see
[CCXV99][XVC00]). The congestion in control plane can be solved by packet queuing in electronic buffers of the controller (e.g., see [KA04]).
Burst losses can be also due to early burst payload arrivals. This effect arises if an
effective processing delay the control packet undergoes in the controller is larger than
a delay budget given by the offset time; in such case the burst is lost. The effective
processing delay is determined by the queuing delay and processing time of control
packet as well as the switch setup time. While control packets experience variable
queuing delays, depending on the congestion situation, the effective processing delays
vary as well. As a result the determination of appropriate delay budget and setup of
offset times that would prevent burst losses is not a trivial task. Notice that excessive
over-provisioning of offset times is undesired in OBS networks since it results in
extended burst delays and puts constraints on the application of fibre delay elements
in E-OBS.
Although, there are some studies that consider the impact of congestion in control plane on OBS node/network performance (see e.g, [WZV02], [BD07]), still, few
of them address the problem of sufficient offset time provisioning. In [BD06] an
initial discussion on some factors which constitute the processing delay budget is
provided. In [KCK04] a control packets scheduling algorithm reducing the effect of
insufficient offsets is proposed. Finally, in [CCXV99] an M/M/1 queuing model is
used to compute an approximation for the complementary distribution of the control
packet processing delay. Since the results presented in these works are very preliminary the study has to be continued.
In order to address thoroughly the problem of sufficient offset time provisioning
49
Chapter 5. Modelling of E-OBS control plane
50
control channels
Control plane
m
m
m
m
Client network
Data plane
ingress node
data channels
m
electronic controller
egress node
core node
all-optical switch
Figure 5.1: Data and control networks of OBS.
the operation in control network has to be analyzed. In particular one has to build a
queuing model of control plane taking into account actual system parameters.
In this Chapter we provide a discussion on several factors that have impact on the
control plane operation. Moreover, we build two exemplary models of E-OBS control
plane which allow as to estimate the delay budget that have to be provided to the
bursts in order to achieve certain target burst loss probability.
5.2
Modelling of control plane
Before elaborating a model of OBS control plane one has to identify both the modelling objectives and the model impacting factors. In particular, the fidelity of model
depends on the phenomenon one wants to study. Some control-plane stability constraints in OBS (see [WZV02][BD07]) can be obtained with a simple algebra based
on basic system parameters. On the other hand a more complex queuing analysis has
to be applied when elaborating a model which involves time dependencies.
5.2.1
Control plane impacting factors
There are many factors that influence the OBS control plane operation and performance; below we list the main of them.
• Network architecture - depending on the use of either a conventional OBS architecture or an offset time-emulated OBS architecture, or some hybrid solution
Chapter 5. Modelling of E-OBS control plane
a) single-processor controller
input
CCs
buffer
51
b) multi-processor controller
output
CCs
output port 1
CPU
CPU
FWD+RR
RR
input
CCs
CPU
FWD
CC - control channel
CPU - processor
FWD - forwarding
RR - resources reservation
output
CC
...
output port N
CPU
output
CC
RR
Figure 5.2: Exemplary controller architectures.
the offset time may either vary or do not inside the network. As a result the
delay budget of bursts entering the node, in principle, is either variable or fixed.
• Node controller architecture - a simple controller can consist of a single processor unit with a buffer handling all the burst control packets in a centralized
way. More advanced controllers can use distributed, pipelined, and parallelized
operation onto multiple processors (e.g., see Figure 5.2). Such architectures
speed-up the processing of control packets.
• Functions and algorithms - the main functions performed by the controller
processors are: forwarding of burst control packets, resources reservation (with
contention resolution and QoS functions) for incoming burst payloads, and configuration of the switching matrix. These functions may be realized with algorithms of different complexity and performance. The algorithm implementation
can be either memory-based, where the processing time depends on the seeking
time in the memory map, or combinatorial, where the processing time is constant. Both selection and implementation of algorithms influence the service
time distribution of the controller.
• Processing technologies - several alternatives exist for the processor implementation, starting from relatively slow processors of general purpose, through the
field programmable gate arrays (FPGA) and network processors (NP), to the
fastest but also the less flexible application-specific integrated circuits (ASIC)
(e.g., see [BBE+ 05]). The first three technologies allow for both memory-based
and combinatorial algorithm implementations, while the ASIC may be limited
only to combinatorial solutions.
• Queuing discipline - either simple first-in, first-out (FIFO) or more advanced
disciplines, for instance with ordering the burst control packets according to
their offsets, can be used in the buffers.
• Data plane-related parameters - the number of both node input/output ports
and data wavelengths has an impact on the amount of burst control traffic
delivered to the controller.
Chapter 5. Modelling of E-OBS control plane
52
packets reneged
from the system
•residence
arrival process of
control packets
Ga
time > delay budget
server
Gs
packets
successfully
processed
Figure 5.3: General OBS control-plane queuing model.
• Characteristics of burst control traffic - the arrival process of burst control packets depends on the burst traffic load, the burst assembly algorithm, in particular on the distribution of both the payload and the control packet lengths, the
number of control channels, and the transmission rates in both control and data
channels.
5.2.2
A queuing model of OBS switch controller
In general, OBS control network is a network of node controllers connected by control
channels. Each controller can be seen as a queuing system. There is some burst control
traffic offered to the controller. The arrival process of control packets is closely related
to the arrival process of data bursts; therefore according to [IA01] it can be modelled
as a Poisson process.
Construction of an accurate queuing model of node controller may be a difficult,
if not impossible, task. The controller service time distribution largely depends on
its features (as discussed before). In particular, the controller architecture could be
represented as a queuing network of buffer-processor systems; some approximation
techniques like for instance a two-moment analysis [Whi83] could be applied here.
The operation of OBS controller can be seen as a queuing with reneging [Boc05].
In particular, a burst control packet, when accepted to the queue, leaves the system
non-served if its delay budget τ is lower than the effective processing delay R , or (in
other words) the residence time (see Figure 5.3). The delay budget is equal to actual
offset-time of the burst. In a well-designed system this offset should be long enough in
order to reduce the probability of burst losses due to their reneging, P = p{R > τ }.
5.3
5.3.1
E-OBS controller with a single processor
Queuing models
We concern on a simple controller with one processor unit and one FIFO buffer, which
handle all the burst control packets arriving to the node. The processing times of the
Chapter 5. Modelling of E-OBS control plane
•a)
M/M/1 with reneging queue
l
¥
53
•b)
M/D/1/K queue
l
1/Tp
P = p{R > t }
•Queue
1/Tp
K
P = p{R > t }
size: infinite
•Queue
•(the
packets not served in time t are
reneged from the system)
size: K = ét / Tpù
•max R = KTp (all the packets accepted
are surely served in time shorter then t)
Figure 5.4: Queuing models: a) M/M/1 with reneging, b) M/D/1/K.
Packet loss probability
Delay budget
Ã
M/M/1 with
reneging
P =
τ (ρ−1)/Tp
(1−ρ)e
1−ρeτ (ρ−1)/Tp
τ=
P ≈
ρ
ρ−1
Ã
M/D/1/K
√ 2τ /Tp − ρ
√
2− ρ
ρ
(ρ−1)
√ τ /Tp +1− ρ
√
2
2− ρ
!
P
ln 1−ρ+P
ρ
τ≈
Tp
√
P
ln 1−ρ+P
−ln(ρ) (2− ρ)
ρ
2 ln(ρ)
!
+ 1 Tp
−1
Table 5.1: Performance of queuing models.
processor are either exponentially distributed (EXP) or deterministic (DET), with
the mean denoted as Tp . We focus on the E-OBS control operation, thus the delay
budget τ of all the bursts entering the node is constant.
Having such a scenario, for each processing time distribution we can consider a
different queuing model, respectively:
• for EXP - M/M/1 queue with reneging (see Figure 5.4a), where all control
packets are accepted to the queue; a packet is lost if period τ expires before the
packet is served.
• for DET - M/D/1/K queue without reneging (see Figure 5.4b), where
control packets are accepted to the queue only if there is free space; when
accepted
h iall those packets are served. The system (queue and server) capacity
K = Tτp guarantees that all the packets entering the queue are served before
period τ . Notice that K gives an upper bound on the packet loss probability of
a M/D/1 queue with reneging.
λ c [1/s]
Chapter 5. Modelling of E-OBS control plane
54
8
10
Stability lines
(λ cTp=1)
7
10
Tp =200ns
6
m
Tp =1µs
5
m
Tp =10µs
10
10
k=4
k=16
k=64
m
T =10µs
4
10
p
m
Tp =1µs
3
T =200ns
10
p
3
4
10
10
5
10
10
6
7
10
mean data burst length, Lb [bytes]
Figure 5.5: Intensity of control packet arrival.
A burst loss probability function P (τ ) and its inverse form τ (P ) are presented in
Table 5.1. We use a fine approximation of M/D/1/K queue which was proposed by
[SC05], whilst we have exact results for M/M/1 queue with reneging [Bar57]. In the
notation, ρ is the processor load (ρ = λTp , where λ is the intensity of control packet
arrival). In the case of M/D/1/K queue we will consider τ to be a multiple of Tp
(τ = KTp ).
5.3.2
Results
The E-OBS node under study has N = 4 input/output ports. The transmission bit
rate of data channel is rb = 10Gbps. We consider fast switch operation with the
switching time TS = 1µs. The analyzed mean processing times are Tp = {10µs,
1µs, 200ns} (as in [BBE+ 05]). We assume the number of control channels is high
enough to carry entire control traffic and to have the packet contention effect in a
control channel negligible.
Control-plane stability
In Figure 5.5 we present the intensity of control packet arrival λc in the function of
average data burst length Lb for the systems with different number of data wavelengths k per port. The burst traffic load ρb is such that the target burst loss probability in data plane PT d = 10−4 ; with the Erlang B-loss formula (see 7.1) we find
it equal to ρb = {0.33, 0.49, 0.62} per wavelength, respectively for the system with
Chapter 5. Modelling of E-OBS control plane
55
0
control packet loss probability, P
10
ρ=0.9, DET
ρ=0.9, EXP
ρ=0.99, DET
ρ=0.99, EXP
-2
P =10-6
10
Tc
ρ=0.99
Target loss probability P =10-6
Tc
-4
10
ρ=0.9
-6
10
0
200
400
600
800
1000
normalized delay budget, tτ/T
p
Figure 5.6: Loss probability of control packets.
k = {16, 32, 64} wavelengths. As we can observe the intensity of packet arrival increases with the number of wavelengths and is inversely proportional to the burst
length.
Moreover for different processing times Tp we plot the boundary λc = 1/Tp of
the control-plane stability constraint ρ = λc Tp < 1 (see [BD07]). Taking this into
account, for each pair of k and Tp we can find the minimum average burst length
which assure the stability of controller operation. Note that with shorter Tp (what
means faster processor operation) this limit can be lowered.
Control-plane loss
We study the impact of delay budget τ (normalized to the processing time Tp ) on
the loss probability P of control packets, for the system with different processor
(controller) load ρ = {0.9, 0.99} and processing time distribution (EXP or DET). As
we can observe in Figure 5.6, P decreases if either τ increases or ρ decreases. With
deterministic processing times we need smaller τ to achieve a certain level of packet
loss probability than in case of exponentially distributed processing times; however,
this difference is reduced with lower ρ. The dotted line delimit a minimum τ which
guarantees a target loss probability in the control plane PT c = 10−6 ; for instance for
EXP and ρ = 0.9 such τ is equal to about 100 times of Tp .
Chapter 5. Modelling of E-OBS control plane
56
l /T =62.6
b
4
10
p
3
m
delay budget, τ [µs]
10
2
10
m
Tp =10µs,DET
1
10
m
T =10µs,EXP
p
m
T =1µs,DET
p
m
Tp =1µs,EXP
0
10
Tp =200ns,DET
Tp =200ns,EXP
Stability asymptote, lb/Tp=Nkρb
stab. asympt.
-1
10
1
2
10
10
10
normalized mean burst duration, l /T
b
3
p
Figure 5.7: Delay budget vs. normalized mean burst duration.
L b ≈14.4kB
35
m
delay budget, τ [µs]
30
L b ≈77kB
m
Tp=1µs
Tp=200ns
m
τ=25µs
25
20
Stability
asymptots
15
10
Tp =1µs,DET
m
m
T =1µs,EXP
p
5
T =200ns,DET
p
T =200ns,EXP
p
0 3
10
4
10
5
10
10
mean burst length, Lb [bytes]
Figure 5.8: Delay budget vs. average burst length.
6
Chapter 5. Modelling of E-OBS control plane
57
Delay budget vs. burst size
In Figure ?? we investigate the impact of normalized mean burst duration lb /Tp on
delay budget τ in the system with k = 32 (and the total number of data wavelengths
N k = 128), different Tp , and target loss probabilities PT d = 10−4 (ρb = 0.49) and
PT c = 10−6 . We can see that if lb /Tp approaches the stability asymptote (N kρb =
62.6) we have τ → ∞ for all curves.
Finally, in Figure 5.8 we plot a reference (dotted) line τ = 25µs corresponding to
the offset provided by a feasible fiber delay coil (see [AST+ 06]). With such target τ we
can find a lower bound on average burst length Lb which preserves the system performance. In particular it is about 100kbytes under moderate, deterministic processing
times (Tp = 1µs), while in the case of fast processing (Tp = 200ns) the limiting value
of average burst length is very close to the one determined by the stability constraint.
5.4
Summary
In this Chapter we address the problem of congestion in the control plane of OBS
network. In order to approach this issue a queuing model of control plane operation
is studied. Since several factors have an impact on the OBS control-plane operation
the elaboration of such model may be a difficult task.
We give some preliminary results for an exemplary E-OBS system with a single processor performing in the node controller. Depending on the distribution of
processing times we model such system either as M/M/1 queue with reneging or as
M/D/1/K queue without reneging. The obtained results show that by appropriate
setup of the minimum mean burst length the congestion in control plane can be effectively limited. Moreover for the analyzed system with moderate processing times we
show that a feasible fibre delay element can both effectively provide the offset times
and concurrently preserve the system performance.
PART III
Quality of service provisioning
Chapter 6
QoS provisioning in OBS networks
The problem of data loss is not uncommon in packet-switching networks. As the
network, or even some of its links and nodes, becomes congested, router buffers fill
and start to drop packets. Another cause can be the changes of routes as a result
of inoperative network links. For non-real-time applications, such as file transfer and
e-mail, packet loss is not critical as long as packet protocols provide retransmission to
recover dropped packets. However, in the case of real-time information, for instance in
voice, video, telemedicine applications, packets must arrive within a relatively narrow
time window to be useful to reconstruct the multimedia signal. Retransmission in
this case would add extensive delay to the reconstruction and would cause clipping
or unintelligible speech as well as discontinuous picture. Packets lost means lost of
some information for these services. Discussed questions led to the introduction of
different quality of service (QoS) classes.
This Chapter addresses the problem of quality of service (QoS) provisioning in
OBS networks. The lack of optical memories results in quite complicated operation
of OBS networks, especially, in case when one wants to guarantee a certain level of
service quality. Indeed, quality demanding applications like for instance real-time
voice or video transmissions need for additional QoS differentiation mechanisms in
order to preserve them from low priority data traffic. In this context the burst blocking
probability metric is perhaps of the highest importance in OBS networks.
QoS differentiation can be provided either with respect to forwarding performance
(e.g., the burst loss rate), or with respect to service availability. In the former case,
certain quality guarantees are expected during a normal, fault-less operation. The
latter case concerns QoS-enhanced protection mechanisms in the resilience problem.
In this thesis we focus on QoS differentiation strategies with respect to the forwarding
performance.
61
Chapter 6. QoS provisioning in OBS networks
6.1
6.1.1
62
Basic concepts of QoS in OBS networks
QoS metrics
Effective QoS provisioning in OBS engages both the definition of specific QoS classes
to be given for higher level applications and some dedicated mechanisms in order
to provide such classes. In general, each class can be characterized by a specific
statistical traffic profile and has to satisfy distinct QoS requirements. In particular,
the requirements concern to ensure a certain upper bounds on end-to-end delay, delay
variation (also called the jitter) and burst loss probability.
The delays arise mostly due to the propagation delay in fibre links, the introduced
offset time, edge node processing (i.e., burst assembly) and optical FDL buffering.
The first two factors can be easily limited by properly setting up the maximum hop
distance allowed for the routing algorithm. Also the delay produced in the edge node
can be imposed by a proper timer-based burst assembly strategy. Finally the optical
buffering, which in fact has limited application in OBS, introduces relatively small
delays. Since there are many factors that influence the end-to-end data delay in OBS
networks the problem of jitter is more complicated and needs a special treatment.
This topic, however, is out of the scope of this thesis.
In a well-designed OBS network the data losses should arise only due to resources
(wavelength) unavailability in a fibre link. The probability of burst blocking in the
link strongly depends on several factors, among other things on the implemented contention resolution mechanisms, burst traffic characteristics, network routing, traffic
offered to the network and relative class load. Since a joint relation between these
factors is usually very complex in formulation the control of burst losses may be quite
awkward in buffer-less OBS networks.
6.1.2
Absolute vs. relative QoS guarantees
There can be distinguished two basic models of QoS provisioning in OBS networks,
namely relative QoS and an absolute QoS. In the former case, the performance of a
class is defined with respect to other classes, for instance it is guaranteed that the
loss probability of bursts belonging to HP class is lower than the loss probability
of bursts belonging to LP class. In the latter case, an absolute performance metric
of quality as for example the maximal acceptable level of burst losses is defined
for a class. The performance of given class in relative QoS model usually depends
on traffic characteristics of the other classes, whilst the absolute QoS model aims
at irrelative quality provisioning. The absolute QoS model requires more complex
implementations in order to achieve desired levels of quality in a wide range of traffic
conditions while at the same time to preserve high output link utilization.
Absolute QoS guarantees are expected by upper level applications. The lack of
optical memories, however, complicates the implementation of absolute QoS model
in OBS networks, comparing for instance to electrical data networks. For this reason
the most of QoS mechanisms considered for OBS networks, basically, offer relative
QoS guarantees.
Chapter 6. QoS provisioning in OBS networks
6.1.3
63
QoS in connection-oriented and connection-less OBS
The problem of QoS guarantees in connection-oriented OBS networks (i.e., with a
two-way signalling) is similar to the one existing in dynamic wavelength-switched
networks. In particular it concerns providing low establishment delays and low connection blocking probabilities, especially for HP connection requests. The establishment delay specifically is critical in such networks. The reason is that the bursts
have to wait in the electrical buffers of the edge nodes until the termination of the
connection establishment process. This may produce the buffer overflow and, as a
consequence, data losses. When the connection is established there is no data loss
inside the network and the transmission delay is only due to the optical signal propagation delay. In this context the connection-oriented OBS operation can provide
absolute quality guarantees.
On the contrary, the one-way signalling model needs for additional support in QoS
provisioning in order to preserve HP traffic from LP traffic during both the resource
reservation process and the burst transmission.
In this thesis we focus on QoS guarantees in one-way signalling OBS networks.
6.2
Categories of QoS mechanisms in OBS networks with one-way signalling
In this section we provide a general classification of QoS mechanisms considered for
OBS networks. In most cases, the contention resolution-based QoS mechanisms have
a similar application in both OBS and OPS networks. Moreover, OBS possesses its
inherent characteristics like for instance the use of pre-retransmission offsets and the
ability to operate with different signalling modes. Such capabilities enable implementation of other QoS schemes that are particular only for OBS networks.
In general, several components can contribute to QoS provisioning in OBS networks with one-way signalling (see Figure 6.1). They are related to the control plane
operation, through signalling and routing functions, and to the data plane operation
both in edge nodes and in core nodes.
6.2.1
Control plane-related mechanisms
Two mechanisms involving control plane operation can provide service differentiation.
On one hand, a hybrid signalling protocol (e.g., see [MGK+ 04]) that consists of a cooperation of two-way and one-way signalling modes can support absolute QoS. In
this scenario the establishment of end-to-end transmission paths, by the two-way
signalling, can provide the guarantees such as no losses and negligible delays inside
the network, while the unreserved resources can be used to transmit the best-effort
burst traffic, with the one-way signalling.
On the other hand, similarly like it was proposed in OPS networks (e.g., see
[ZCC+ 04][YMY01]), a routing function can support QoS provisioning. In particular,
a properly designed routing protocol may minimize the path lengths for delay-sensitive
Chapter 6. QoS provisioning in OBS networks
64
Mechanisms for QoS
provisioning
Data plane
Control plane
Routing
Signalling
Edge node
Core node
Burst dropping
schemes
Offset-time
differentiation
Varying assembly
parameters
Preemptive
dropping
Threshold
dropping
Scheduling differentiation
of control packets
Intentional burst
dropping
Figure 6.1: Categories of QoS mechanisms in OBS networks.
applications, and even preserve the selection of overloaded parts of the network for
loss-sensitive ones, for instance thanks to a deflection routing operation (as e.g., in
[LKSG03][KG03a][LYH+ 06]).
6.2.2
Edge-based mechanisms
Edge nodes are responsible for the burst assembly process so that the incoming client
packets are aggregated into bursts in the electronic buffers, according to their class and
destination. Solutions like [VJ03], where bursts are unaware class assembled, involve
additional complexity and they are used only in particular cases, e.g., together with
a burst segmentation mechanism.
Then QoS can be achieved in the following ways:
• Offset Time Differentiation [YQ98], which is probably the most addressed QoS
technique in OBS networks. The idea here is to assign an extra offset-time to
high priority bursts, which results in an earlier reservation, in order to favor
them while the resources reservation is performed (see Figure 6.2a). The offset
time differentiation mechanism allows to achieve absolute isolation between HP
and LP classes, i.e., no HP class burst is blocked by a LP class burst. To
achieve such feature, however, the extra offset has to be as large, at least, as
a few average LP burst durations. The main advantage of this technique is
its simplicity; it reduces the loss probability of HP bursts by their postponed
transmission from the edge node and no differentiation mechanism is needed in
the core nodes. The disadvantages are both the sensitivity of HP class to burst
length characteristics [DG01] and extended pre-transmission delay that may not
be tolerated by some time-constrained applications. Another problem of the
offset-time differentiation mechanism in C-OBS networks is the multiplication
of effective classes due to the offset variation [DG01] which may impair the
class isolation. In order to reduce this effect a processing offset, which gives a
Chapter 6. QoS provisioning in OBS networks
a) Offset Time Differentiation
Control
HP
65
b) Burst Preemption (BP)
LP
Partial preemption
Full preemption
In1
LP burst
In1
HP burst
In1
HP burst
Standard OT
In2
HP burst
Standard + Extra OT
Blocking
HP
Out
PT
Out
In2
LP burst
In2
LP burst
Out
LP burst
Succesful
resources
reservation
HP burst
HP burst
HP burst
time
time
c) Burst Dropping with Wavelength threshold (BD-W)
Wavelengths
HP bursts
l1
access for HP bursts
l2
LP: Low Priority
HP: High Priority
OT: Offset-Time
PT: Processing Time
In: Input
Out: Output
l: Wavelength
Control: Control Channel
Threshold
l3
access for LP bursts
l4
time
LP bursts
Figure 6.2: Selected QoS mechanisms in OBS networks.
margin to the processing and switching operation in core nodes, should be small
enough.
• Varying burst assembly parameters like, e.g., preset timers or burst lengths. In
particular, the packets belonging to HP class can be aggregated with shorter
burst assembly timers than LP packets [BS04]. In this way the latency experienced by HP traffic can be minimized. Also, in the networks with FDL-buffering
and void-filling capabilities shorter bursts can improve the BLP performance of
a HP class [KCSSP05]. Designing of a burst assembly algorithm is a delicate
task as long as the resulting traffic characteristics may influence overall network
performance.
Another QoS function at the edge node is the classification of traffic with assignation of specific attributes to the bursts like, e.g., labels, or priorities. These attributes
are carried by burst control packets with the purpose of their further discrimination
and processing in core nodes.
6.2.3
Core-based mechanisms
First of all, QoS provisioning in core nodes takes place when resolving the contention
problem and is achieved with the assistance of a burst drooping technique. The contention resolution usually is supported by some mechanism(s) like wavelength conversion, FDL buffering or deflection routing. The following burst dropping techniques
have been proposed for QoS differentiation in OBS:
• Preemptive dropping, which in case of the burst conflict overwrites the resources
reserved for a lower priority burst by a higher priority one; the preempted,
Chapter 6. QoS provisioning in OBS networks
66
LP burst is discarded (see Figure 6.2b). Several variations of the preemption
mechanism can be found in the literature and both relative and absolute QoS
models are supported by this technique (e.g., see [OS06][YJJ03]). In general the
preemption can be either full or partial. The full preemption concerns the entire
LP burst reservation [KA03] while the partial preemption overwrites only the
overlapping part of LP reservation [VJ03]. The partial preemption allows for
more efficient resources utilization comparing to the full preemptive scheme. Its
drawback, however, is additional complexity in the burst assembly process since
this technique requires additional information about the burst data segments
to be carried and processed in core nodes.
• Threshold-based dropping, which provides more resources, like wavelengths or
buffers, to HP bursts than to LP ones according to a certain threshold parameter (see Figure 6.2c). If the resources occupation is above the threshold, the
LP bursts are discarded and the HP bursts are accepted until there are some
resources available. Likewise the OPS networks, in which some threshold-based
algorithms have been proposed to be used for the wavelength and FDL buffer
assignment problem [CCRZ04], similar solutions can be applied easily in OBS
networks [ZVJC04].
• Intentional bursts dropping, which maintains the performance objectives of the
higher priority bursts on certain levels by intentional dropping the lower priority
bursts. As a discarding scheme, a random early detection (RED) technique can
be used [ZVJC04]. The intentional burst dropping may be classified as an
absolute QoS technique.
Another group of mechanisms, which support QoS provisioning in core nodes, is
based on queuing and scheduling management of burst control packets that arrive to
the node controller. These mechanims make use of the observation that by a proper
ordering of burst control packets some reservation requests can be processed earlier;
as a result, there is more chances to encounter free transmission resources. Some of
proposed burst control packet scheduling mechanisms are taken directly from wellstudied electrical packet networks. For instance, in [WR04] the burst control packets
are processed according to their priorities, while in [KA04] a fair packet queuing
algorithm, which regulates access to the reservation manager for different classes of
services, is applied. A disadvantage of priority scheduling techniques in OBS networks
is extended burst delay. Indeed an additional offset time has to be introduced in order
to give enough time for gathering the burst control packets, for the purpose of their
prioritized scheduling in the controller.
Chapter 6. QoS provisioning in OBS networks
67
QoS
mechanism
Implemented
QoS model
Supported
QoS
parameter
Advantages
Disadvantages
Hybrid
signalling
absolute
delay /
burst losses
- absolute end-to-end loss
and delay guarantees for
HP
- lower statistical multiplexing
gain, inefficient usage of bandwidth (less resources available
for LP traffic)
QoS routing
absolute
(delays)
relative
(burst losses)
delay /
burst losses
- introduces QoS guarantees on network level
- controlling burst losses may
be challenging (need the knowledge about network state)
relative
burst losses
- simple, soft operation
- no need for any differentiation mechanism in core
nodes
- sensitivity of HP class to burst
length characteristics
- extended pre-transmission delay
absolute
(delays)
relative
(burst losses)
delay /
burst losses
- assembly parameters can
be easily setup
- the resulting traffic characteristics may influence network
performance
burst losses
- fine class isolation
- improved link utilization
in scheme with a partial
preemption
- absolute QoS can be
achieved with a probabilistic preemptive scheme
- overbooking of resources in
consecutive nodes (in case of
successful preemption)
- additional complexity involved in the burst assembly
process in case of partial preemption
Offset-time
differentiation
Varying burst
assembly
parameters
Preemptive
dropping
relative /
absolute
Thresholdbased dropping
relative
burst losses
Intentional
burst dropping
absolute
burst losses
Scheduling
differentiation
of control
packets
relative
burst losses
- can be easily implemented
- can provide absolute QoS
- priority queuing in electrical buffers is a feasible and
well studied technique
- the efficiency of bandwidth
usage strongly depends on
threshold adaptability to traffic
changes
- the link utilization may suffer
- complex implementation
- extended delay (need for
longer queuing windows and so
larger offset times to perform
effectively)
Table 6.1: Characteristics of QoS mechanisms in OBS networks with one-way signalling
In Table 6.1 we summarize the main features of discussed QoS mechanisms.
Chapter 7
Performance of QoS mechanisms in
E-OBS
When examining the literature one can find several proposals of QoS mechanisms in
OBS networks (see Chapter 6). Usually it is difficult to compare their performance
as each mechanism is evaluated in a specific node/network scenario. Nevertheless,
a few works provide comparative performance results of selected QoS mechanisms.
For instance, in [ZLW+ 04] Zhang analyzes different QoS provisioning scenarios, which
apply either a wavelength threshold-based or an intentional burst dropping principle,
with the purpose of absolute quality guarantees. In [VJ03] Vokkarane compares the
performance of different QoS schemes with a burst segmentation approach applied.
Also the evaluation of different optical packet-dropping techniques in an OPS network
scenario is provided in [OS06].
In this Chapter we make an extension to these studies. In particular, we confront
the performance of a frequently referenced offset time differentiation (OTD) mechanism with two burst-dropping techniques, namely, with a burst-preemptive dropping
(BP) and a wavelength threshold-based dropping (BD-W). In general, all these mechanisms aim at differentiation of burst loss rates in OBS networks that operate with a
one-way signalling. We confront the performance of mechanisms in the E-OBS architecture and under a connection-less UDP traffic scenario. Two classes of traffic are
considered, namely, a high priority (HP) class and a low priority (LP) class.
7.1
7.1.1
Overview
QoS scenario details
All QoS mechanisms are evaluated in a unified single core-node scenario (see Figure
7.1). There is a number of edge nodes which generate some HP class and LP class
burst traffic pattern. The burst traffic is handled in the core node according to a given
resources reservation and burst drooping policy. The performance of QoS mechanisms
is evaluated using the output burst traffic characteristics at the node output link.
Our core node implements the E-OBS architecture, i.e., the offset times are intro69
Chapter 7. Performance of QoS mechanisms in E-OBS
Edge Node (EN1)
Classifier
Core Node
BCP generator
(w/wo priority
labeling)
BCP
HP class
BCP
Burst
assembler
IP packets
Resources reservation
(w/wo QoS-aware
burst drooping)
Offset management
(w/wo QoS diff.)
LP class
Scheduler
Switch
fabric
Burst
Burst
...
LP: Low Priority
HP: High Priority
BCP: Burst Control Packet
70
Fiber
Span
Results
BLPHP, BLPLP
BLPoverall
throughput
ENN
Figure 7.1: Evaluated QoS network scenario.
duced by means of additional fibre delay coils. On the contrary to C-OBS, there is
no additional offset, except an optional extra QoS offset, introduced in the edge node
between a burst control packet and its data payload. Thus we avoid the impact of
variable-offsets on the scheduling operation [LQXX04] and we can get deeper insight
into the mechanisms behavior. Since the scheduling operation affects all the mechanisms equally we can expect that their relative performance will be also preserved in
C-OBS.
We focus on a (nowadays) technologically feasible OBS core node (e.g., [GWL+ 05]
[AST+ 06]), which operates with relatively low number of input ports and wavelengths
but with fast, sub-microsecond switching matrix and short burst durations.
We consider that in the core node the burst scheduler uses a void filling-based algorithm. Our implementation of the algorithm searches for a wavelength that minimizes
the time gap which is produced between currently and previously scheduled bursts.
We assume that the searching procedure is performed according to a round-robin rule,
i.e. each time it starts from the less-indexed wavelength.
The implementation of QoS mechanisms is as following:
• The size of extra offset time assigned to HP bursts in the offset time differentiation mechanism is equal to 4 times of the average LP burst duration. According
to [YQD01] it assures quasi-absolute class isolation.
• We consider a simple full-preemptive scheme when implementing the burst preemption mechanism. Particularly, each HP burst is allowed to preempt at most
one LP burst if no free wavelength is available. The preemption concerns a LP
burst the dropping of which minimizes the gap produced between the preempting HP burst and the rest of burst reservations.
• The wavelength threshold-based burst dropping mechanism performs according
to a restricted approach (e.g., see [OS06]). In particular, the threshold value
specifies the maximum number of wavelengths that can be simultaneously occupied by LP bursts. On the contrary, the HP bursts are allowed to access the
whole pool of wavelengths. The threshold selection problem is discussed in the
next subsection.
Chapter 7. Performance of QoS mechanisms in E-OBS
71
In both scenarios, the preemptive burst dropping and the wavelength thresholdbased dropping, there is a traffic classification function implemented in the edge node,
which assignes priorities to the bursts.
The metrics we evaluate are:
• burst loss probability (BLP), which corresponds to the amount of data bursts
lost in the core node among all the data bursts generated in the edge nodes,
and
• effective data throughput, later called the throughput, which represents the percentage of data volume served with respect to overall data volume offered to
the core node.
The burst loss probability is obtained with respect to both HP and LP class traffic
as well as to the overall traffic.
In this study we are interested in quantitative comparison of QoS mechanisms
more than in the system design or dimensioning. Therefore some of the simulation
parameters are setup so that to obtain the evaluation results, especially in case of the
HP class performance, in reasonable simulation times.
7.1.2
Simulation scenario
We set up an event-driven simulation environment to evaluate the performance of
QoS mechanisms. The simulator imitates an E-OBS core node with no FDL buffering
capability, full connectivity, and full wavelength conversion. It has 4×4 input/output
ports and 8 data wavelengths per port (if not specified otherwise), each one operating
at 10Gbps. The switching times are neglected in the analysis.
The traffic is uniformly distributed between all input and output ports. In most
simulations the offered traffic load per input wavelength is ρ = 0.8Erlang (i.e., each
wavelength is occupied in 80%) and the percentage of HP bursts over the overall burst
traffic, also called HP class relative load αHP , is equal to 30%.
The burst length is normally distributed with the mean burst duration L = 32µs
and the standard deviation σ = 2 · 10−6 . In further discussion we express the burst
lengths in bytes and we neglect the guard bands. Thus the mean burst duration
L corresponds to 40kbytes of data (at 10Gbps rate). The burst arrival times are
normally distributed with the mean that depends on the offered traffic load and the
standard deviation σ = 5 · 10−6 .
All the simulation results have 99% level of confidence.
7.2
Threshold selection in BD-W mechanism
One of designing problems in any threshold-based mechanism is how to specify the
threshold value. The selection of threshold can be supported by an appropriate analysis of the mechanism performance. An analytical queuing model of the wavelength
threshold-based dropping (BD-W) mechanism was presented in [OS06]. Based on
Chapter 7. Performance of QoS mechanisms in E-OBS
a)
b)
c)
d)
72
Figure 7.2: Performance of BD-W mechanism (c = 8), a) HP class BLP, b) LP class
BLP, c) throughput, d) threshold value guaranteeing BLPHP ≤ 10−4 .
this model, we present some performance characteristics of BD-W mechanism as well
as we present a method of threshold selection. In our analysis we consider a system
with 16 wavelengths, 0.8 traffic load and exponentially distributed burst arrivals and
burst lengths. We use Tw to denote the threshold value.
Let Erl(·) be the Erlang’s loss formula:
" c
#−1
Ac X Ai
Erl(A, c) =
(7.1)
c! i=0 i!
where A is the offered load and c is the number of wavelengths.
In Figure 7.2 we present exemplary results of HP and LP class burst loss probabilities and the throughput. We can see that the performance of BD-W mechanism
depends on both HP relative traffic load (αHP ) and threshold value (Tw ). For given
αHP one can adjust BLPHP by a proper selection of the threshold, however at the
cost of effective throughput. On one hand, the minimum BLPHP is delimited by
b1 = Erl(αHP ρ, c), and achieved with Tw = 0. In this case the LP class traffic is not
served. On the other hand, the maximum BLPHP is equal to b2 = Erl(ρ, c), and
obtained for Tw = c. Such a scenario means that no class differentiation is provided
Chapter 7. Performance of QoS mechanisms in E-OBS
73
by the mechanism.
∗
Assuming that some target burst loss probability BLPHP
is higher than b1 , we
∗
∗
∗
can find some threshold Tw that complies BLPHP (Tw ) ≤ BLPHP and maximize the
throughput. Figure 7.2d presents such evaluation as a function of ρ and αHP for
∗
BLPHP
= 10−4 and c = 8 wavelengths.
7.3
7.3.1
Performance results
Burst loss probability and throughput
In our study both the offset time differentiation (OTD) and the burst preemption
(BP) can be characterized by absolute class isolation. In particular the extra offset
time of OTD mechanism assures that the contention of HP bursts is only due to the
other HP burst reservations. Therefore, if we assume the exponentially distributed
burst arrivals and lengths, the burst loss probability of HP class can be modelled
with formula (7.1), and it corresponds to Erl(αHP ρ, c). Similarly, the BP mechanism
allows to preempt any LP reservation by a HP one, and the loss of a HP burst
happens only if all the wavelengths are occupied by HP reservations. As a result, the
loss probability of HP bursts can be expressed as Erl(αHP ρ, c) again. Note that both
schemes can successfully transmit LP bursts if either there are free wavelengths, not
occupied by any earlier HP reservations in the case of OTD mechanism, or they are
not preempted by HP bursts in the case of BP mechanism.
As it was already discussed, the BD-W mechanism achieves the best HP class
performance if there is no threshold established (Tw = 0); i.e. only HP bursts are
transmitted at the output port. In such case, the HP class burst loss probability of
BD-W mechanism is the same as of both OTD and BP mechanisms. However, the
throughput of BD-W mechanism is very low since none LP burst can be served. As
Figure 7.2 shows, by increasing the threshold we can improve the throughput, but at
the cost of worsen HP class performance.
In Figure 7.3 we provide comparative simulation results for the scenario described
in the previous section, namely for ρ = 0.8 and αHP = 30%. The evaluation is
performed for different number of data wavelengths (c) in a link. We establish Tw ,
the wavelength threshold of BD-W mechanism, equal to 50% of c so that LP class
bursts can access at most the half of all the available wavelengths simultaneously.
As we can see in Figure 7.3a, by increasing the number of wavelengths in a link
we improve the effectiveness of QoS differentiation. The improvement of BLPHP in
both OTD and BP mechanism can be so high as, for instance, of three orders of
magnitude when having 16 instead of 8 wavelengths. Also, we can see that BD-W
offers the worst HP class performance among the evaluated mechanisms.
Comparing BLPLP , the overall BLP and the effective throughput (Figure 7.3b-d),
although the shape of OTD and BP performance characteristics is similar, still, the
results are in favor of the BP mechanism (see the next subsection for more details).
Regarding the BD-W mechanism, we can see that is exhibits very poor performance,
which hardly depends on the number of wavelengths. The reason is that BD-W
b)
1,E+00
Offset-Time Differentiation
Burst Preemption
Burst Dropping w ith Wavelength threshold
1,E-01
1,E-02
1,E-03
1,E-04
1,E-05
1,E-06
1,E-07
4
8
16
32
LP class Burst Loss Probability
a)
HP class Burst Loss Probability
Chapter 7. Performance of QoS mechanisms in E-OBS
1,E+01
Offset-Time Differentiation
Burst Preemption
Burst Dropping w ith Wavelength threshold
1,E+00
1,E-01
1,E-02
4
64
d)
1,E+01
120%
Offset-Time Differentiation
Burst Preemption
32
64
Burst Dropping w ith Wavelength threshold
Burst Dropping w ith Wavelength threshold
1,E+00
16
Offset-Time Differentiation
Burst Preemption
110%
100%
Throughput
Overall Burst Loss Probability
8
Number of wavelengths
Number of wavelengths
c)
74
1,E-01
1,E-02
90%
80%
70%
60%
1,E-03
50%
4
8
16
32
Number of wavelengths
64
4
8
16
32
64
Number of wavelengths
Figure 7.3: Performance of QoS mechanism vs. link dimensioning (ρ = 0.8, αHP =
30%), a) HP class BLP, b) LP class BLP, c) overall BLP, d) effective data throughput.
mechanism has effectively fewer wavelengths available for the burst transmissions in
the output link than the other mechanisms. Indeed, BD-W provides only 50% of
wavelengths for LP class bursts, while at the same time it attempts to serve the same
amount of offered burst traffic. As a result, both the LP class burst loss performance
and the throughput are seriously deteriorated.
Although the FDL buffering, in principle, is hardly considered in OBS networks,
the utilization of short data bursts may enable its application in the contention resolution problem. The application of FDL buffering should improve the utilization of
link resources, and thus the node throughput, as well as it should decrease the loss
probability of bursts belonging to each priority class (e.g., see [BNH+ 03]).
7.3.2
Burst preemption vs. offset time differentiation
The results of HP class burst loss performance presented in Figure 7.3a and Figure
7.4a confirm the correctness of argumentation provided in the previous section. In
particular, we can see that BLPHP of both OTD and BP is comparable under any
traffic load conditions (Figure 7.4a) as well as with different link dimensioning (Figure
7.3a).
On the other hand, in Figure 7.4b we can see that LP traffic is served more
efficiently in BP mechanism than in OTD mechanism. The explanation to this fact
can be found in [LQXX04], where it is shown that the scheduling operation may
be impaired by the variation of offset-times, the feature which is inherent to OTD
Chapter 7. Performance of QoS mechanisms in E-OBS
a)
75
b)
HP class
LP class
1,E+00
1,E+00
Burst Loss Probability
Burst Loss Probability
1,E-01
1,E-02
1,E-03
1,E-04
Offset-Time Differentiation
Offset-Time Differentiation
Burst Preemption
Burst Preemption
1,E-05
1,E-01
0,2
0,4
0,6
0,8
1
0
0,2
HP class relative load
0,4
0,6
0,8
1
HP class relative load
Figure 7.4: Burst loss probabilities vs. HP class relative load in OTD and BP mechanisms (ρ = 0.8, c = 8), a) HP class, b) LP class.
load = 0.5
a)
load = 0.8
b)
89%
98%
88%
Throughput
Throughput
87%
97%
86%
85%
84%
83%
Offset-Time Differentiation
Offset-Time Differentiation
82%
Burst Preemption
Burst Preemption
81%
96%
0
0,2
0,4
0,6
HP class relative load
0,8
1
0
0,2
0,4
0,6
0,8
1
HP class relative load
Figure 7.5: Effective throughout vs. HP class relative load in OTD and BP mechanisms, with overall traffic load: a) 0.5, b) 0.8.
Chapter 7. Performance of QoS mechanisms in E-OBS
76
mechanism. As Figure 7.5 shows, the use of variable offsets impairs the effective
data throughput of OTD mechanism, especially, if the classes are equally loaded.
Comparing Figure 7.5a and 7.5b we can see that the aggravation of throughput is
more serious in highly loaded nodes.
There is some deterioration of the effective data throughput in the BP mechanism.
It is a result of the preemptive operation which allows dropping of a LP burst even if
it is being transmitted. Since the front (already transmitted) part of this burst has
used some link bandwidth, it also increased the effective traffic load offered to the
link; hence, the burst blocking probability increases.
7.4
Summary
In this Chapter we study performance of the most addressed mechanisms providing
relative QoS differentiation in OBS networks with one-way signalling. In particular,
we show that a burst preemptive mechanism concurrently achieves efficient resources
utilization and offers highly-effective QoS differentiation. An offset time differentiation mechanism, which is frequently invoked in the literature, also provides high
HP class performance, however, its scheduling efficiency, and so the throughput, is
aggravated by the variation of offset times. Finally, a wavelength threshold-based
mechanism can be characterized by the poorest overall performance that significantly
depends on the threshold value. The application of this mechanism may be reasonable only in the networks with a large number of wavelengths in the link, where the
wavelength threshold parameter would be relatively high (in order to serve the LP
traffic efficiently) and it could adapt according to the traffic changes. Although the
performance of QoS mechanisms has been evaluated in a single node scenario, still,
we can expect their similar behavior in a network scenario.
Chapter 8
Effective burst preemption in
E-OBS
Burst preemption is one of the most frequently applied techniques to provide burst
differentiation in OBS networks. This technique can be used both in QoS provisioning
(see Chapter 6) as well as in OBS routing [CZZ04][LY06]. In the previous Chapter we
show that in a single node scenario a preemption-based burst dropping mechanism
is characterized by high overall performance. Analyzing the mechanism behaviour in
a network scenario, however, we encounter the problem of so called phantom bursts.
The phantom burst is a burst, which was dropped in a node, but its control packet
is still travelling through the network reserving the transmission resources. Such
situation takes place for instance in the conventional OBS networks with a burst
preemption mechanism applied. This effect may lead to the wastage of transmission
resources and a superfluous control processing effort. In fact either the overbooked
resources in downstream nodes are wasted or an additional signalling procedure should
be carried out in order to release them. Moreover the control packets belonging to
the phantom bursts burden the controllers of switching nodes unnecessarily.
In this Chapter we estimate the amount of additional signalling necessary to release the overbooked resources in a single buffer-less OBS node. Then we present a
novel control mechanism which efficiently applies the burst preemption technique in
an E-OBS node without the resources overbooking. Analytical and simulation results
prove the effectiveness of our proposal.
8.1
Preemption rate in a buffer-less OBS node
In order to estimate the amount of additional signalling, in this section we calculate
a preemption rate metric (R) that expresses the amount of preempted bursts over all
successfully transmitted bursts at the node output port. Since each preemption would
involve a signalling message to release the resources on the ongoing path such metric
corresponds well to the signalling overhead produced in a node. In this analysis we
consider a full-preemptive scheme, which means that the preemption concerns the
entire burst reservation. Moreover we assume the exponentially distributed burst
77
Chapter 8. Effective Burst Preemption in E-OBS networks
78
arrivals.
(np)
(p)
Let npreempt be the number of successful preemptions, nlost HP and nlost HP be the
number of HP bursts lost in a non-preemptive scenario (without burst preemption)
and a preemptive scenario (with full burst preemption) respectively, nin HP be the
number of incoming HP bursts, nin be the total number of incoming bursts and nout
be the total number of bursts transmitted at the output in a given period of time.
Since each preemption means the acceptance of a HP burst instead of a LP burst,
npreempt can be also interpreted as a difference between all the HP bursts lost in the
non-preemptive scenario and the HP bursts lost in the preemptive scenario:
(np)
(p)
npreempt = nlost HP − nlost HP .
(8.1)
Obviously:
(np)
(np)
(8.2)
(p)
(p)
(8.3)
nlost HP = nin HP · BHP ,
nlost HP = nin HP · BHP ,
(np)
(p)
where BHP and BHP are the HP burst loss probabilities in the non-preemptive scenario and the preemptive scenario.
From the previous equations we obtain:
³
´
³
´
(np)
(p)
(np)
(p)
npreempt = nin HP · BHP − BHP = αHP · nin · BHP − BHP ,
(8.4)
where αHP is the HP class load ratio.
Than the preemption rate is equal to:
R=
npreempt
=
nout
³
´
(np)
(p)
αHP · nin · BHP − BHP
nin · (1 − B (p) )
.
(8.5)
Notice that the overall burst loss probability in the preemptive scenario (B (p) )
(np)
and the HP class burst loss probability in the non-preemptive scenario (BHP ) are the
(p)
same. Moreover, BHP depends only on the HP class load due to the absolute class
isolation principle. Finally, assuming exponentially distributed burst arrivals, we can
use (7.1) to calculate burst loss probabilities. Therefore, by proper substitution we
obtain the following estimation of the preemption rate in a node:
R=
αHP [Erl (ρ, c) − Erl (αHP ρ, c)]
1 − Erl (ρ, c)
(8.6)
where ρ, αHP , c, respectively, are the overall traffic load, HP class load ratio and the
number of wavelengths in a link, and Erl(·) is given by (7.1).
The numerator of the formula 8.6 indicates the reduction of burst losses of the
HP class due to the preemption with respect to the non-preemptive scenario, while
the denominator conditions the preemption only to those bursts that are successfully
allocated.
Chapter 8. Effective Burst Preemption in E-OBS networks
b)
HP class load = 30%
HP class load = 50%
1,E+00
1,E+00
1,E-01
1,E-01
Preemption Rate (R)
Preemption Rate (R)
a)
1,E-02
1,E-03
load=0.5 (analytical)
load=0.8 (analytical)
1,E-04
79
1,E-02
1,E-03
load=0.5 (analytical)
load=0.8 (analytical)
1,E-04
load=0.5 (simulation)
load=0.5 (simulation)
load=0.8 (simulation)
load=0.8 (simulation)
1,E-05
1,E-05
4
8
12
16
20
24
Number of wavelengths
28
32
4
8
12
16
20
24
28
32
Number of wavelengths
Figure 8.1: Percentage of additional signalling necessary to release preempted burst
at each node, with HP class load: a) 30%, b) 50%.
Figure 8.1 presents both analytical and simulation results of the preemption rate
in a single node scenario. As we can see, R increases significantly in the systems
with lower number of wavelengths as well as at higher traffic loads. A small disparity
between analytical and simulation results comes from the fact that the simulated
bursts are stream-like arranged in a data channel (bursts do not overlap each other)
and their arrivals are not more exponentially distributed.
As we already said, R corresponds to the percentage of additional signalling required (at each node) to release the preempted bursts. If such a signalling procedure
is not provided, there is waste of transmission resources due to the phantom reservations in all the nodes on the ongoing routing paths. In large networks, of high number
of nodes, the problem might be intensified since all nodes undergo a similar effect.
A particular attention should be paid to preemptive-based routing mechanisms.
For instance, a deflection routing mechanism proposed in [CZZ04] as well as a burst
cloning mechanism proposed in [LY06] assume that the bursts carried over alternative
(duplicate) paths can be preempted by the bursts carried over primary paths. In
such a scenario the amount of preempted bursts would be really high as long as
both ρ and αHP are high. As a consequence the phantom burst reservations, which
occupy the transmission resources unnecessarily, decrease the effectiveness of routing
mechanisms.
Evaluation of the phantom burst effect in a network scenario is out of the scope
of this thesis.
8.2
Preemption Window (PW) mechanism
In this section we propose a control mechanism that overcomes the problem of resource overbooking due to the burst-preemptive operation. For the purpose of our
mechanism we define a time window in which the preemption of LP burst is allowed.
Chapter 8. Effective Burst Preemption in E-OBS networks
CC
LP
t1
HP
t2
T
DC1
D
80
HP
t3
LP burst (1)
t0
preemption allowed
(t2<t0)
DC2
HP burst (2)
D
DC3
preemption not allowed
(t3>t0)
HP burst (3)
D
D: (one-hop) offset time
T: preemption offset
time
HP/LP: high/low priority
CC: control channel
DC: data channel
Figure 8.2: Principles of the preemption window mechanism.
The preemption window (PW) mechanism expands look-ahead processing window
techniques to the burst preemption context.
8.2.1
Principles
In the PW mechanism a control packet is delivered to the switch controller with some
extra offset, besides the processing offset time. This additional offset constitutes a
preemptive window T during which the controller can preempt the lower priority
reservation by the higher priority reservation.
An important rule of the PW mechanism is that the control packet, after its
processing, is waiting in the memory of the controller until T expires and only then
it can be sent to the next node (if the burst has not been preempted) or dropped (in
case of successful preemption). After the control packet is sent, the preemption of its
burst is not allowed in the node. Thanks to these rules each control packet has its
corresponding data burst existing in the network (no phantom bursts are present),
and there is no need for any signaling procedure to be carried out in order to release
the resources on the outgoing path in case of successful burst preemption.
Figure 8.2 shows an illustrative example of the PW mechanism. In this example,
a preemption of the LP burst (1) can be performed only by the HP burst (2) since
the control packet of the latter burst arrives in preemptive window T . On the other
hand, the HP burst (3) is not allowed to preempt the LP burst (1) because its control
packet arrives out of window T .
The main advantage of the proposed mechanism is the lack of signalling overhead
in case the preemption occurs. Indeed when the control packet reserves resources it
already knows that its burst will arrive to the node. There is no resources overbooking
in downstream nodes, and so, there is no need to release them. It should be pointed
Chapter 8. Effective Burst Preemption in E-OBS networks
Control packet
Time
CP
DC
81
preemptive window
LP burst (payload)
T
dp
di
standard D
Dp
dp: effective queuing and processing delay
di: idle waiting time after the processing
D: offset time introduced by FDC
Dp: additional preemptive offset
D in PW
Figure 8.3: The length of preemptive window in PW mechanism.
out that the PW mechanism can work with any, either full or partial, burst preemption
principle.
The preemption offset can be provided in both C-OBS and E-OBS architectures.
In the former the edge node adds an additional offset, which accounts the preemption
windows in all the nodes of the routing path. A disadvantage of this solution is the
increase of variation of offset times, which may further intensify the unfairness in
access to transmission resources (see Chapter 4). For this reason we consider the PW
mechanism more appropriate to be used with E-OBS architectures.
8.2.2
The length of preemptive window
Preemptive window T begins after the end of processing of the burst control packet
and lasts until the arrival of its payload (see Figure 8.3). In further discussion, for
simplicity, we assume that the payload incorporates a guard band for the switching
operation.
Period T can be calculated as:
T = ∆ − δp
(8.7)
where ∆ is the offset introduced by inlet FDC in E-OBS node, and δp is the effective
queuing and processing delay of control packet.
Since δp is variable (see Chapter 5 for more details) period T is variable as well. In
the simplest case, T corresponds to the idle waiting time period δl after the processing
of control packet. In order to increase this period, the FDC can add some additional
preemptive offset ∆p . In this case T could be also expressed as:
T = δi + ∆p
(8.8)
In the context of burst differentiation, the value of T becomes an important tradeoff between high burst delay (too large preemptive window) and ineffective burst
preemption (too short preemptive window). Scope of the following sections is to
determine the minimum value of T that provides optimal blocking probability.
Chapter 8. Effective Burst Preemption in E-OBS networks
82
lLP
lHP
HP
lHP
plHP
m
lLP
lLP
0
LP
m
(1-p)lHP
Figure 8.4: Markov chain representing a single-wavelength model of PW mechanism.
8.3
A single-wavelength model of PW mechanism
In this section, we analyze the blocking probabilities of the two classes of bursts,
namely a high priority (HP) and a low priority (LP) class, in a single channel system
when a full-burst preemption mechanism with PW principle is applied.
According to [IA01] we assume a Poisson process for the HP and LP burst arrivals
with rates λHP and λLP respectively; the overall arrival rate will be λ = λHP + λLP .
Let t be an i.i.d. exponential random variable which denotes the burst inter-arrival
time. Also, let l denotes the burst duration, which follows an exponential distribution
with mean value 1/µ. Although, we assume the same distribution for both classes,
still, in further analysis we use lLP in order to emphasize that we mean the duration
of an LP burst.
Let denote the possible wavelength states as: Q0 for free wavelength, QLP if
occupied by an LP burst, and QHP if occupied by an HP burst. A corresponding
Markov chain, which determines the states’ transition, is presented in Figure 8.4.
The equlibrium equations are the following:
pλHP PLP + λHP P0 = µPHP ,
(8.9)
λLP P0 = (µ + pλHP )PLP ,
(8.10)
P0 + PHP + PLP = 1,
(8.11)
and
where P0 , PLP , and PHP denote the probabilities of being in states Q0 , QLP , and
QHP respectively.
By solving this system of equations we can easy determine the state probabilities
Chapter 8. Effective Burst Preemption in E-OBS networks
T
a)
lHP
lLP
T
Successful preemption
LP
HP
t
Port1
HP burst
Port2
Preemption
Window
CPout
HP burst
b)
T
Preemption fail
lHP
lLP
T
HP
LP
Port1
LP burst
Input
CPin
Output
HP
Port1
t + T < lLP + T
t>T
Input
LP burst
t
HP burst
Port2
CPout
Preemption
Window
LP
LP burst
Port1
Output
t + T < lLP + T
t<T
CPin
83
Figure 8.5: (a) Successful preemption and (b) preemption fail cases; the processing
times are neglected for simplicity, T is the duration of the Preemption Window, lLP
and lHP are the durations of the LP and HP bursts respectively, t is the arrival time
of the HP control packet.
P0 =
µ
,
λ+µ
(8.12)
PLP =
λLP µ
,
(µ + p λHP ) (λ + µ)
(8.13)
PHP =
λHP (µ + p λ)
.
(µ + p λHP ) (λ + µ)
(8.14)
Consequently, the burst blocking probability of LP bursts (BLP ) and HP bursts
(BHP ) can be calculated as
BLP = PLP + PHP + p
λHP
PLP,
λLP
BHP = PHP + (1 − p)PLP ,
(8.15)
(8.16)
where p is the probability of successful preemption (referred to as P (Y )) with respect
to all attempts of preemption (referred to as P (A)).
Figure 8.5 helps to discriminate successful and failed preemptions. In particular,
in Figure 8.5a the preemption of an LP burst can take place since the control packet
Chapter 8. Effective Burst Preemption in E-OBS networks
84
of the HP burst arrives in the preemption window. The HP burst in Figure 8.5b is
not allowed to preempt the LP burst because its control packet arrives out of the
preemption window.
Considering the i.i.d. exponential random variable, the probability of successful
preemption P (Y ) can be calculated as
P (Y ) = P ((tHP < lLP ) ∩ (tHP < T ))
Z T Z ∞
=
µ e−µx λHP e−λHP y dx dy
0
y
λHP
=
(1 − e−(µ+λHP )T ).
λHP + µ
(8.17)
The probability P (A) is equal to
P (A) = P (tHP < lLP ) + P ((tHP < lLP ) ∩ (tHP > T ))
∞
X
P (tHP < lLP )i ,
(8.18)
i=1
and it represents a first HP burst arrival when the wavelength is occupied by an LP
burst (tHP < lLP ) plus all further HP arrivals (the summation) in case of preemption
fail (tHP < lLP and tHP > T ).
By solving (8.18) we obtain
∞
X
λHP
λHP
λHP
−(µ+λHP )T
P (A) =
+
e
(
)i
µ + λHP
µ + λHP
µ
+
λ
HP
i=1
=
λHP
λHP −(µ+λHP )T
(1 +
e
).
µ + λHP
µ
(8.19)
The probability of successful preemption p is therefore
p=
P (Y )
1 − e−(µ+λHP )T
=
.
P (A)
e−(µ+λHP )T
1 + λHP
µ
(8.20)
Taking into account (8.15), (8.16) and (8.20)
BLP =
λHP µ (1 − e−(λHP +µ)T )
λ
+
,
λ+µ
(λHP + µ)(λ + µ)
(8.21)
BHP =
λ
λLP µ (1 − e−(λHP +µ)T )
−
.
λ+µ
(λHP + µ)(λ + µ)
(8.22)
Chapter 8. Effective Burst Preemption in E-OBS networks
8.3.1
85
Some remarks
Having the burst blocking probabilities for LP class (8.21) and HP class (8.22), we
can derive overall burst blocking probability Boverall that is given by
Boverall =
λHP
λLP
λHP + λLP
λ
BHP +
BLP =
=
.
λHP + λLP
λHP + λLP
λHP + λLP + µ
λ+µ
(8.23)
As we could expect, the obtained result corresponds to the Erlang loss formula
(7.1). Indeed, the PW mechanism does not impair the total blocking probability
and even in the case of preemption, when a LP burst is replaced by a HP one, the
number of lost bursts is preserved. Also, notice that the formula does not involve T
parameter, contrary to both BHP and BLP .
Now, let’s check the burst blocking probabilities under the boundary conditions.
For T = 0 we obtain
BLP = BHP =
λHP + λLP
,
λHP + λLP + µ
(8.24)
which is also equal to Boverall . It is clear, because since T = 0 there is no preemption
(NP) and the mechanism performs as a poor scheduling mechanism without QoS
differentiation.
Now, for T → ∞ we have
lim BLP =
T →∞
λHP + λLP
λHP µ
+
,
λHP + λLP + µ
(λHP + µ)(λHP + λLP + µ)
(8.25)
and
lim BHP =
T →∞
λHP + λLP
λLP µ
λHP
−
=
.
λHP + λLP + µ
(λHP + µ)(λHP + λLP + µ)
λHP + µ
(8.26)
We can see that for T → ∞ both BLP and BHP exponentially approach their
asymptotes defined by constant functions of λHP , λLP and µ parameters. In particular
the second asymptote for BHP could be also derived from the Erlang loss formula with
only the HP traffic taken into account. The explanation is that since T → ∞ the
length of LP bursts would be almost surely less than T and therefore an HP burst
can be blocked only by another HP burst. In this case the mechanism behaves like a
classical preemption algorithm (CP), where an HP burst can always preempt an LP
burst.
Figure 8.6 presents the characteristics of the discussed model (PW model), validated by simulation results (PW sim). Notice, that the x-axis on the graph is normalized by the mean burst duration (1/µ) and α is the HP traffic ratio.
The PW model gives a glance on the mechanism’s behavior in a single-wavelength
system. To complete the study in the next section we provide simulation results of
PW mechanism in a multi-wavelength scenario.
Chapter 8. Effective Burst Preemption in E-OBS networks
86
0.65
Burst Blocking Probability
0.6
0.55
0.5
0.45
PW model (HP class)
PW model (LP class)
CP (HP class)
NP (HP class)
PW sim (HP class)
PW sim (LP class)
PW sim (Overall)
0.4
0.35
0.3
0.25
0.2
0
0.5
1
1.5
2
TmT·µ
Figure 8.6: Simulation vs. modeling results (ρ = 0.8, α = 0.3, µ = 2).
8.4
Computer simulation of PW mechanism
We use an event-driven simulator to evaluate the performance of a full-burst preemptive mechanism with PW applied. We look for an effective offset, introduced by
means of the inlet FDC, which is a trade-off between offering high performance and
minimizing the delay. We assume there are two classes of traffic, namely a HP class
and a LP class.
8.4.1
Simulation scenario
We evaluate the PW mechanism in a single buffer-less OBS node with full wavelength
conversion, 4 × 4 input/output ports and operating at 10Gbps. The LAUC scheduling
with full-burst preemption is applied. For the purpose of simplicity the processing and
switching times are set to 0. The traffic is uniformly distributed between all ports.
We consider two traffic models: a general Exponential and a specific Gaussian burst
length and inter-arrival time distributions; the latter represents the traffic generated
by a hybrid time-length burstifier. Both models use 40kbytes (32µs) as mean bursts
length; for Gaussian model we set up the standard deviation to 2µs, and minimum
and maximum burst lengths to 4kbytes and 4M bytes, respectively. The mean burst
inter-arrival times depends on the offered load ρ. The HP burst traffic ratio over
overall one is denoted as α. All the simulation results have 99% level of confidence.
Chapter 8. Effective Burst Preemption in E-OBS networks
87
Figure 8.7: Burst blocking probability as a function of T comparing Gaussian and
Exponential traffic models (α = 30%, ρ = 0.8, W = 16).
8.4.2
Numerical results
In Figure 8.7, we firstly compare the Classical Preemption (CP) with our Preemption Window (PW) solution as a function of the delay T . When T = 0, there is
no possibility of preemption and PW performs as a simple scheduling without burst
differentiation. When T increases, HP (LP) burst blocking probability decreases
(increases) and approximates to an asymptote, which corresponds to the results obtained with CP. In case of Gaussian traffic, PW quickly reaches the CP performance
(T larger than 30µs), while worse results are obtained with Exponential one (T larger
than 60µs). This is because the former generates a concentration of burst durations
more closed to the length of the fibre delay coil than latter; it has to be underlined
that this Gaussian traffic model can be easily obtained well tuning the time/length
thresholds of the burstifier [YLC+ 04].
As Fig. 8.8 shows, burst blocking probability would be further reduced in the
systems with more wavelengths. We can discern that for T ≥ 30µs (6km of FDC)
and W ≥ 16 wavelengths, HP burst blocking probability is less than 10−6 .
In Figure 8.9, we analyze the blocking probability as a function of the offered load
ρ and of the percentage of HP burst traffic load α. The T window is fixed to 10µs
(2km) and 32 wavelengths are considered. We can observe that PW achieves very
low HP burst blocking probabilities, e.g. 10−5 at ρ = 0.65 and α = 40%. Again, PW
behaves better when the burst generation follows the Gaussian model.
Chapter 8. Effective Burst Preemption in E-OBS networks
88
Figure 8.8: Burst blocking probability as a function of T and of W (α = 30%, ρ = 0.8,
Gaussian traffic model).
Figure 8.9: Burst blocking probability as a function of ρ comparing Gaussian and
Exponential traffic models and different α (T = 10µs and W = 32).
Chapter 8. Effective Burst Preemption in E-OBS networks
89
8λs, 1FDL (BLPHP)
8λs, 2FDL (BLP
Burst Loss Probability (BLP)
0
10
)
HP
8λs, 4FDL (BLPHP)
8λs, 4FDL (BLP )
LP
8λs, 4FDL (BLP
-2
)
Overall
32λs, 1FDL (BLPLP)
10
-4
10
-6
10
0
0.5
Tm
T·µ
1
1.5
Figure 8.10: Burst blocking probability as a function of T (normalized to 1/µ) for
different W and FDL buffer size (α = 25%).
8.4.3
PW and FDL buffering
Now we consider that the switching node is enhanced with a feed-back FDL buffer
(e.g., as in [HCA98]). The feed-back node architecture allows us to preempt any LP
burst even, at the moment, it is transmitted through the FDL buffer. In fact, when
a preemption occurs we know that the LP burst has not reached the output port yet,
thanks to the PW rule. Hence, we can easily block this burst in the switching matrix
(after its looping through the FDL), and thus make impossible its further propagation
towards an output link. It have to be noticed that the preemption of a burst which
is transmitted through a feed-forward FDL buffer might result in the propagation of
a part of optical signal, which has not been blocked by the switching matrix. Since
this useless part of the burst would reach the next node it could cause false optical
signal detections and therefore additional information such as jam sequence might be
required.
In our analysis we assume that the feed-back buffer emulate N output feed-forward
buffers, each one operating with 8 optical channels, where N is equal to the number
of output ports. The number of delay lines is between 1 and 4 depending on the
simulation. The provided delays are linearly increasing with a basic delay unit equal
to 32µs, which corresponds to the mean burst duration.
In Figure 8.10 we show the results of BLP for different buffer size and number
of wavelengths as a function of T (normalized to mean burst duration 1/µ). We see
that even with one FDL used there is no significant gain in the performance when
Chapter 8. Effective Burst Preemption in E-OBS networks
90
increasing T . It is due to the fact that the buffer itself introduces some variable
preemption window and therefore no extra preemptive offset in the inlet FDC is
necessary. This also explains why, even with T equal to 0, the results of BLPHP are
much lower than BLPLP and BLPOverall . Therefore the length of input FDC and its
consequent delay produced in the node can be reduced. Notice that the controller still
has to postpone the transmission of control packets in order to send them together
with the corresponding bursts.
Finally, we can observe that the application of FDLs decreases blocking probability
of LP bursts, In particular, in the system with 32 wavelengths (λs) and just with only
1 FDL the BLPLP can be below 10−4 in a node.
8.5
Summary
The high overall performance of a preemption-based burst dropping mechanism designates it to be a suitable burst differentiation mechanism in OBS networks. However,
in this thesis we are concerned on relative quality guarantees, the preemption technique can be extended to absolute QoS provisioning as well. Such a study can be found
e.g., in [OS06]. There the superiority of a packet preemptive technique over other
packet differentiation techniques, including an intentional packet dropping technique,
is demonstrated again.
In OBS the main drawback of a burst preemption mechanism is the overbooking
of resources in case of a successful burst preemption.
In this Chapter we propose a dedicated control mechanism, called the preemption window mechanism, which preserves from resources overbooking. PW allows
for preemption of a low priority burst only in a specific preemptive window period.
Although, our mechanism can be applied in C-OBS, still, it benefits more from the
E-OBS control architecture.
Both modeling and simulation results show that the PW mechanism achieves the
same performance of the conventional preemptive scheme. The obtained values of the
preemption window show the feasibility of its application; e.g., a fibre of about 6km
is enough when a Gaussian distributed burst traffic model is applied.
Furthermore, the PW mechanism can be used with any other preemptive technique
like burst segmentation.
Finally, in the scenarios with FDL buffering there is no need for extra preemption
offset in order to obtain QoS differentiation since it is provided by the FDL buffer
itself.
PART IV
Routing
Chapter 9
Routing in OBS networks
OBS architectures with no buffering capabilities are sensitive to burst congestion.
The existence of a few highly congested links may seriously aggravate the network
throughput (e.g., see [KHCSP05]). A burst loss probability (BLP) which adequately
represents the congestion state of entire network is the primary metric of interest in
OBS networks.
The congestion can be reduced either by appropriate network dimensioning or by
proper network routing. The dimensioning approach fits the node and link capacities
according to the matrix of actual traffic load demands and after such optimization it
needs only either a simple shortest path algorithm or a similar mechanism (e.g., see
[KG03b][GKZM05]). Some parts of such network, however, may encounter the congestion problem if the traffic demands change. On the contrary, the routing approach
introduces some operational complexity since it often needs advanced mechanisms
with signalling protocols involved. Nevertheless, the advantage is that it facilely
adapts to the changes in traffic demands. Since both presented solutions complete
rather than substitute each other, an OBS network should be designed with both
a proper link dimensioning and an adequate routing strategy operating inside the
network.
In this part of the thesis we address the problem of network routing in the context
of burst loss performance and congestion reduction.
9.1
9.1.1
Introduction
Routing terminology
Routing algorithms can be grouped into two major classes: non-adaptive and adaptive (see Figure 9.1) [Tan88]. Non-adaptive, also called static, ones do not base their
routing decisions on measurements or estimates of the current traffic and topology,
whereas adaptive, or dynamic, ones do.
In static routing the choice of the route to use to get from node A to node B
is computed in advance, off-line, and downloaded to the nodes when the network
is booted. Thus, routing variables do not change during the time. The simplest
93
Chapter 9. Routing in OBS networks
94
b) isolated
a) centralized
c) distributed
d) sing-path
e) multi-path
f) alternative
g) source
primary path
alternative
path
node
link
route (path)
processing of routing
information exchange
collision (congestion)
Figure 9.1: Routing algorithms.
technique for static routing is based on a shortest path routing algorithm, where the
routing objective is to find a routing path of minimum length. The path length in
the shortest path routing can be calculated in several ways: the number of hops and
the geographic distance are the easiest metrics; for the former we will alternatively
use the term shortest hop routing.
On the other hand, adaptive algorithms, attempt to change their routing decisions
to reflect changes in topology and the current traffic. Adaptive algorithms can be
further divided into three families, which differ in the information they use, namely:
• centralized (or global) - a single entity uses information collected from the
entire network in an attempt to make optimal decisions,
• isolated (or local) - a local algorithm runs separately on each node, which only
uses information available there, such as e.g., output link congestion,
• distributed - uses a mixture of global and local information.
Chapter 9. Routing in OBS networks
95
So far we have tacitly assumed that there is a single path between any pair of nodes
and that all traffic between them should use it. Such routing approach is usually called
single-path routing. In many networks, there are several paths between pairs of
nodes that are almost equally good. Better performance can frequently be obtained by
splitting the traffic over several paths, to reduce the load on each of the communication
links. The technique of using multiple routes is called multi-path routing. An
advantage of multi-path routing over single path routing is the possibility of sending
different classes of traffic over different paths. It can also be used to improve the
reliability of the network, in particular, if the routing tables contain disjoint routes.
Alternative routing, often referred to as deflection routing, is a special case of
multi-path routing. Later we distinguish alternative routing as a technique where all
the traffic is sent over a primary routing path. In case the primary path is unavailable
for soma period of time a secondary, alternative path is selected.
Another distinction in routing algorithms can be with respect to the place where
the routing decision is taken. Whilst most of routing algorithms can perform in each
node, in source routing only the source makes most or all of the routing decisions.
Thus, with source routing the entire path to the destination is known to the sender
and is included when sending data. Source routing allows a source to directly manage
network performance by forcing data to travel over one path to prevent congestion
on another.
9.1.2
Reactive and proactive burst loss reduction techniques
To reduce burst losses reactive and proactive techniques are applied in the network
[TVJ03].
• Reactive techniques, e.g., wavelength conversion, FDL buffering, deflection
routing, attempt to resolve burst contentions rather than avoid the contentions.
Usually, they are based on a local information at the node.
• Proactive techniques, reduce the number of burst contentions, by policing
the traffic at the source (buffering or dropping data), or by routing traffic in a
way that the congestion in the network is minimized. A proactive policing at
the source may be controlled by feedback information that indicates congestion
in the network.
Most routing-based proactive techniques involve two stages; route calculation
and route selection.
The route calculation can be divided into two categories, namely static and dynamic. In a static-route calculation, one or more routes are calculated ahead of time,
based on some static metric, such are physical distance or number of hops. For instance, paths can be computed using Dijkstra’s shortest-path algorithm. In general
these static techniques are suitable when the traffic if fairly steady; however they may
suffer if traffic is fluctuating over time.
In dynamic route calculation techniques the routes are computed periodically
based on certain transient (dynamic) traffic information such as link congestion or
Chapter 9. Routing in OBS networks
96
number of contentions. Route computation can be performed either centrally in a
predestinated node or distributively in individual network nodes.
The information necessary to make the route computation can be obtained in
two ways, namely probe-based or broadcast-based. In the probe-based approach,
the source node sends a probe message into the network. The core nodes respond
to the probe and return necessary information to the source. A particular case of
probe messaging could be a feed-back notification about successful (ACK) or failed
(NACK) burst transmission. In the broadcast approach, the core node is responsible
for transmitting relevant congestion information periodically to all the edge nodes.
The probe can either be sent once for every connection request or periodically based
on some interval. The second option is preferable in OBS networks since the duration
of data bursts is usually short. In order to reduce the control traffic in the broadband approach, the feedback information can be sent only if there is a change in the
congestion status of a link from the previous value.
Once the routes are computed, one of the routes is selected for the data transmission. In single-path routing, the route-selection stage is omitted. In a static
route-selection, the traffic is splitted so that its fixed fraction is sent on each of the
routing paths. Dynamic route-selection policies are based on feedback information,
like in dynamic route-calculation techniques. For each route a given cost function is
performed so that the routes are ranked according to their congestion state. Both the
traffic splitting vector and the route ranking techniques should react to link congestion states and adopt accordingly in order to shift some part of traffic to less-loaded
links.
Stabilization is a significant issue in dynamic route calculation and selection. In
particular, multiple sources when reacting to congestion simultaneously, may result
in oscillation between congested and un-congested states on particular links. Hence,
such effect should be avoided in the network.
9.1.3
Hop-by-hop vs. explicit routing
Routing of data through the network can be performed either hop-by-hop, like e.g.,
in connectionless IP networks, or explicitly from source-to-destination, like e.g., in
connection-oriented multi-protocol lambda switching (MPLS) networks.
• In hop-by-hop routing, or datagram-based routing, a control packet contains
the destination address of the burst, based on which layer 3 forwarding (or
routing) is done at every intermediate node.
• In explicit routing, or virtual connection-based routing, a logical connection,
also called label switched path (LSP), is set-up first over an explicit physical
route. Each control packet carries a label (an LSP identifier), based on which
layer 2 forwarding (or switching) is done at every intermediate node. As a
result, all bursts sent on an explicit route will follow the path through to the
destination. The collection of LSPs between various pairs of nodes essentially
forms a virtual network on top of the physical fibre network topology.
Chapter 9. Routing in OBS networks
[WMA02]
[KKK02]
[HLH02a]
[CWXQ03]
[ZVR+ 04]
[VJ02b]
[CZZ04]
[LKSG03]
[KG03a]
[LYH+ 06]
[CEJ05b]
[HAM+ 05]
97
Type
Inform.
Routes
Deflection
Other
S
S
S
S
S
S
S
A
A
A
A
A
I
I
I
I
I
I
I
C
C+D
C
D
D
S
S
S
S
S
S
S
O
O
O
S
S
F
F
F
P
F
F
F
F
T
P
R
T
Q
Q
Q
Q
-
Type: non-adaptive, static (S), adaptive (A)
Information: isolated (I), centralized (C), distributed (D)
Routes: static (S), dynamic, optimized (O)
Deflection: fixed (F), threshold-based (T), probabilistic (P), rank-based (R)
Other: QoS-aware (Q)
Table 9.1: Classification of literature on alternative routing in OBS networks.
Normally, layer 2 forwarding is based on finding an exact match between the
label carried by a packet and a label created during the LSP set-up process and
accordingly, it is faster than layer 3 forwarding. As well, the simplicity of layer 2
forwarding can facilitate traffic engineering (i.e., intentional distribution of traffic
over the network) and end-to-end QoS. These capabilities fit well to both high-speed
processing requirements of node controllers and the need for constrained routing, in
order to preserve from link overloads, of buffer-less OBS architectures. As a result,
the use of labelled optical burst switching (LOBS) has been proposed in [Qia00] as a
natural control and provisioning solution under the MPLS framework.
9.2
9.2.1
State of the art
Alternative routing
A great part of research on routing problem in OBS networks concerns alternative
(or deflection) routing. In alternative routing, when the burst contention occurs, a
deflective mechanism reacts to it and re-routes a blocked burst from the primary to
an alternative route. Deflection routing can be combined with other burst contention
resolution mechanisms (e.g., see [VJ02b][GKS04]).
Routing strategies considered for alternative routing in OBS networks can be
either non-adaptive or adaptive.
In non-adaptive alternative routing both primary and alternative routing paths
Chapter 9. Routing in OBS networks
[TR05]
[PMP07]
[OA05]
[LMC05]
[AdDA07]
[LLGC06]
[TVJ03]
[ACP04]
[GBIQ04]
[IYS05]
[HTM06]
[YR06a]
98
Inform.
Routes
Selection
Selection
Method
Other
C
C
D
D
D
D
D
D
D
D
D
D
O
S
S
S
S
S
S
S
S
S
S
S
P
P
P
P
P
P
R
R
R
R
O
O
H
H
H
O
H
H
H
H
H
Q
-
Information: centralized (C), distributed (D)
Routes: static (S), optimized (O)
Selection: static, probabilistic (P), dynamic, rank-based (R)
Selection Method: optimized (O), heuristic (R)
Other: QoS-aware (Q)
Table 9.2: Classification of literature on adaptive multi-path routing in OBS networks.
are fixed (static), and in most cases calculated with the Dijkstra algorithm. A number
of alternative paths can be given from a node to the destination. Routing decision is
taken in isolation, based only on a local node congestion state information.
Adaptive alternative routing strategies apply a proactive calculation of alternative
paths as well as their dynamic selection. The calculation of alternative paths is performed in an optimized way with the assistance of linear programming formulations.
These methods need for the information about network topology and traffic demands.
In the case of dynamic alternative route selection some heuristics methods are used.
In particular, either threshold-based or path-rank (priority) or probabilistic route selection techniques are applied. Dynamic route selection methods need for distribution
of some link/node state information between respective nodes. Some of alternative
routing strategies support QoS provisioning by routing differentiation with respect to
the quality class.
Table 9.1 summarize the key literature on the alternative routing in OBS networks.
9.2.2
Multi-path routing
Multi-path routing strategies in OBS networks aim in adaptive distribution of traffic over a number of routing paths in order to reduce network congestion. Although,
some proactive optimization techniques can be found [TR05], still Dijkstra’s shortestpath algorithm is the most explored method for pre-calculation of routing paths. In
most cases a small number of disjoint SPs with respect to the number of hops is
Chapter 9. Routing in OBS networks
[HN04]
[ZLW+ 04]
[ZWZ+ 04]
[LY06]
[TR05]
[CMC06]
[OTYC05]
[DPZQ06]
[Bou03]
[HHM05]
[GZ06]
99
Inform.
Method
Inform.
type
Other
C
C
C
C
C
C
C
C
D
D
D
O
O
O
O
O,H
O,H
H
H
H
H
H
T
T
T
T
T
T
B
B
B
F
F
F
F
F
Information: centralized (C), distributed (D)
Method: optimization (O), heuristic (H)
Information type: topology with traffic demands given (T), broadcasted (B)
Other: failure-recovery implemented (F)
Table 9.3: Classification of literature on adaptive single-path routing in OBS networks.
calculated between each source-destination pair of nodes. In OBS multi-path routing, the selection of routing path is performed in the source. The path selection can
be either according to given probability, like in the multi-path routing with traffic
splitting, or according to the path congestion rank. Some authors propose centralized, optimization methods for the calculation of traffic splitting vector whilst the
others apply distributed heuristic methods. A ranking of the less-congested paths
usually is obtained with some distributed heuristics algorithms. In both cases the
distributed methods need for updates about the network state information from intermediate/destination nodes to the source nodes. Such signalling messages can be
either broadcasted or based on some events, like for instance, the burst dropping
event.
Table 9.2 summarize the key literature on the multi-path routing in OBS networks.
9.2.3
Single-path routing
Both non-adaptive (static) or adaptive (dynamic) strategies are considered for singlepath routing in OBS networks. Static routing is usually based on Dijkstra’s shortest
path calculation with respect to the number of hops (e.g., see [YR06b]).
Adaptive single-path routing aims in burst congestion avoidance thanks to a proactive path calculation. The path calculation can be performed either in a centralized
or in a distributed way. Centralized (or pre-planed) routing in OBS, in most cases,
makes use of the optimization theory with (mixed) integer linear programming formulations. In each case it is supposed that a route computation unit has a knowledge
Chapter 9. Routing in OBS networks
100
about network topology and (long-term) traffic demands. On the contrary, distributed routing uses some heuristics. Node state statistics are broadcasted, usually in
a periodical manner, and used to calculate link weights (costs) in respective nodes.
Then a Dijkstra-like calculation is applied in order to find the lowest cost route. Some
of adaptive single-path routing strategies support network resilience by the computation of backup paths.
Table 9.3 summarize the key literature on the adaptive single-path routing in OBS
networks. The table is structured by horizontal lines according the the main criteria.
9.3
Summary
Dijkstra’s shortest path algorithm is the primary routing strategy, frequently explored
in OBS networks. Shortest path routing reduces overall network utilization when
calculated with respect to the number of hops. On the other hand, some links may
be overloaded, while others may be spare, leading to excessive burst losses. Therefore
several both reactive and proactive routing strategies have been proposed with the
objective of the reduction of burst congestion.
First research studies concerned alternative routing with a static route calculation
and selection. Although deflection routing improves network performance under low
traffic conditions [WMA02], still, it may intensify the burst losses under moderate
and high loads [ZVR+ 04]. Indeed the problem of alternative routing in buffer-less
OBS networks is over-utilization of link resources, if an alternative route has more
number of hops than a primary path. Therefore in the next step the objective was
to optimizate the set of alternative routes as well as to introduce some adaptive path
selection techniques (see Table 9.1). Assigning of lower priorities to deflected bursts
with their possible preemption is another important technique, which preserves from
excessive burst losses on primary routes [CZZ04].
Multi-path routing represents another group of routing strategies, which aim in
traffic load balancing in OBS networks (see Table 9.2). Most of the proposals are
based on a static calculation of the set of equally-important routes with Dijkstra’s
algorithm. Then the path selection proceeds adaptively according to some heuristic
or optimized cost function. Both traffic splitting and path ranking techniques are
used in the path selection process.
The issue related to any multi-path routing is the problem of out-of-order burst
arrivals. The burst reordering is common for both multi-path and alternative routing
scenarios, in which the routing paths differ with respect to a physical distance. To
cope with this problem either some dedicated mechanisms (e.g., see [GBIQ04][LMC05]
[PMP07]) or single-path routing have to be used.
Network congestion avoidance in single-path routing is achieved thanks to a proactive route calculation. Since most of the strategies proposed for OBS networks consider a centralized single route calculation some authors study distributed routing
algorithms (see Table 9.3). Both optimization and heuristic methods are used. Moreover, several works address the problem of network resilience and failure recovery.
Chapter 10
Isolated alternative routing
strategies for labelled E-OBS
networks
In this Chapter we propose two isolated alternative routing algorithms. Our objective is to find the algorithm that at the same time can be easily implemented in
a connection-oriented, labelled E-OBS network and improves the overall burst loss
performance. As a reference we use a simple shortest hop routing (SPR) algorithm.
The evaluation is performed in an event-driven simulator environment. All simulation
results have 99% level of confidence.
10.1
Scenario under study
Network architecture
We consider an OBS network with one-way signalling, Horizon resources reservation,
LAUC burst scheduling and E-OBS architecture. The application of E-OBS facilitates routing management. In particular, there is no constraint on the length of an
alternative path as well as the offsets do not have to be computed in source nodes in
advance but they are introduced accordingly in immediate core nodes. Notice that
since the offset time is fixed in corresponding E-OBS nodes there is no need for a void
filling-based burst scheduling algorithm.
Each network node is both an edge node and a core switching node capable of
generating bursts destined to any other nodes. In the analysis we assume that the
source nodes do not buffer the bursts after completing their aggregation. Also, the
nodes are not enhanced with FDL buffers.
Number of data wavelengths c is the same for each link and equal to c = {32, 64},
depending on the scenario. The transmission bitrate of data wavelength is 10Gbps.
101
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
102
SIMPLE
NSFNET
EON
number of nodes N
6
15
28
number of links K
8
23
41
minimum node degree
2
2
2
maximum node degree
4
4
5
average node degree
2.67
3.07
2.928
minimum link length [km]
500
247
218
maximum link length [km]
500
2831
1500
average link length [km]
500
1022
625
network diameter (hops)
3
4
8
Table 10.1: Network topologies
Network topologies
Our routing strategies are evaluated with three logical network topologies (see Figure
10.1):
• the SIMPLE mesh network topology,
• the NSFNET network topology, which represents an American backbone network [Nsf], and
• the EON network topology, which is a pan-European network defined in European COST 266 action [RI03].
The SIMPLE network has 6 nodes and is the smallest network. On the other hand
the EON network (28 nodes) is the largest network. The number of nodes (N ) and
links (K) in the NSFNET network (15 nodes) can be placed in-between. Average
node degree (2K/N ) is approximately the same for both the NSFNET and the EON
networks. Maximum and average link length is significantly larger in the NSFNET
network compared to the other two networks. The NSFNET network contains both
rather short and very long links.
Network diameter, which is the maximum distance between node pairs based on
the number of hops, is a good indicator for the amount of through traffic in network
nodes.
The details on the topologies can be found in Table 10.1.
Traffic model
The traffic is uniformly distributed, i.e., the following matrix of demands T is defined:
T=
cρ
(E − I).
N −1
(10.1)
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
a) SIMPLE network
1
2
3
4
5
6
b) NSFNET network
c) EON network
Figure 10.1: Network topologies; a) SIMPLE, b) NSFNET, and c) EON.
103
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
104
where N is the number of network nodes, c is the number of wavelengths in the
network link, ρ is the traffic load offered to edge node normalized to the link capacity,
E is the unit matrix, I is the identity matrix, and all the matrixes has dimension
N × N . In other words, between each pair of source-destination nodes there is a
traffic offered. The volume of traffic is equal to the amount of traffic load entering
the edge node divided by the number of corresponding destination nodes.
We consider a Poisson arrival process for generating bursts with exponentially
distributed lengths. As several authors already observed it (e.g., see [CEJ05b]), the
length distribution does not have a significant effect on the results in buffer-less OBS
networks.
Route calculation and selection
We assume the routing paths are calculated according to Dijkstra’s shortest hop algorithm. In all studied routing algorithms we consider that there are k pre-established
LSPs between all source-destination pairs of nodes available. The routes are not
necessarily disjoint.
An LSP selection is performed according to a given routing algorithm. Particularly, isolated alternative routing allows to select an LSP, from the set of all available
LSPs, in any network node. We consider per-burst routing decision.
10.2
Algorithms
We propose two isolated alternative routing algorithms, namely, a path excluding
routing (PER) algorithm and a by-pass routing (BPR) algorithm. Each algorithm
performs a deflection of transmitted data burst from a primary to an alternative
routing path if there are no transmission resources available on the primary path.
The routing decision is taken only using local (isolated) output link state information.
Path excluding routing algorithm
In PER algorithm the edge node selects the first available path from the set of paths
to the destination. This selection determines the next hop and excludes from the set
of available paths all those paths that not include this hop in their route. Hence,
from the k original paths, each node removes some paths as long as there remains
only one path.
Figure 10.2a-b show an example. A burst is generated in node A and destinated
to node E. k = 3 paths are setup: the shortest ones are 1. A − D − E, then 2.
A − B − C − E and 3. A − D − F − E. If the first (shortest) path from the list is
congested on its output port, A selects the A−B −C −E path definitely excluding the
other possibilities. This means that the rest of the nodes in the selected path cannot
take other routing decisions. If the output port of A toward D is not congested, both
A − D − E and A − D − F − E are selected while the other is removed. The next
node D will take the path decision in the same way. If the output port of D toward
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
a) Path excluding routing
A
105
b) Path excluding routing
B
C
D
E
A
F
B
C
D
E
F
c) Bypass path routing
A
B
p
By
s
as
th
pa
D
C
By
E
pa
ss
pa
th
F
Figure 10.2: Isolated alternative routing algorithms: a) PER, and b) BPR.
node E is not congested, it chooses the path D − E; otherwise D − F − E is selected.
It is evident that when all output ports are congested, the burst is lost.
Bypass path routing algorithm
In BPR algorithm, for each burst, the source node selects a single path as a function of
the state of its output queues. The route can be modified only when travelling burst
finds a congested link. In this case, the node tries to by-pass it using the shortest
available path to the destination.
Figure 10.2c shows an example of this algorithm behavior. Node A transmits a
packet/burst to node D with destination node E (the path is A − D − E). When
burst arrives to node D, no resources are available to reach node E. Therefore, node
D finds two by-pass paths in its forwarding table: D − C − E, and D − F − E. It
selects the first available one.
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
10.3
106
Results
Our isolated alternative routing algorithms are evaluated in the network scenarios
described in Section 10.1; in particular, each link has c = 32 data wavelengths in
SIMPLE and NSFNET network, and c = 64 wavelengths in EON network. We
consider the scenarios with k = {2, 4, 6} LSPs between each pair of nodes available
in SIMPLE and NSFNET networks, while there can be k = {2, 4, 6, 8, 10} LSPs in
EON network. The offered traffic load ρ is normalized to the link capacity.
In Figures 10.3 and 10.4 we present the impact of the number of available paths
(LSPs) on overall BLP performance under PER and BPR routing strategies respectively.
Firstly, we can see that both PER and BPR outperform SPR under low and
moderate traffic loads in each scenario. Moreover, the efficiency of PER under high
loads (ρ > 1) can be still better than of SPR, whilst BPR has a worsen performance.
These results are consistent with the conclusions presented in [ZVR+ 04]. Particulary,
BPR algorithm does not have any limits on the number of deflections performed and
it can increase the network load, and so the burst blocking, significantly. On the
other hand, the number of deflections in PER is limited, at most, to the number of
available paths k. The network is hardly overloaded in such case.
The next conclusion is that more LSPs improves the network performance. It is
obvious since there are more possibilities to perform the deflection in case of unavailability of resources on primary paths. The improvement of performance can be really
high under BPR strategy and in smaller networks (see Figure 10.4a). BPR with high
number of LSPs available behaves like a hot-potato routing; the burst can be sent
even to the previous node (loops possible). Nevertheless, the selection of the set of
LSPs should be reasonable in order to preserve from the use of too-long paths, as
e.g., in Figure 10.4c), where the performance with k = 10 is worsen than with k = 8.
When comparing the routing strategies we can see that BPR offers better performance than PER (except high-load traffic conditions). Again, it is clear since BPR
has more chances for a successful deflection in every intermediate node.
In Figure 10.5 we investigate the distribution of the number of hops the burst,
which is successfully delivered to the destination, experiences with BPR in SIMPLE
and NSFNET network scenarios. As a reference we provide the similar distribution
obtained with SPR; the maximum number of hops here is 3 and 4 for SIMPLE and
NSFNET respectively. We can see that BPR can increase the length of the burst
routing path significantly, especially, under higher loads (compare Figure 10.5b with
Figure 10.5c). On the other hand, under PER strategy the maximum burst routing
path in the network is limited by the length of the longest LSP (it results from the
behavior of routing algorithm).
10.4
Summary
Isolated alternative routing performs the route selection in consecutive nodes based
on local node state information (e.g., link occupancy, available wavelengths), i.e. each
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
a)
107
SIMPLE, 32 wavelengths
1,E+00
Burst loss probability
1,E-01
1,E-02
1,E-03
1,E-04
SPR (1LSP)
PER (2LSPs)
1,E-05
PER (4LSPs)
PER (6LSPs)
1,E-06
0,4
0,6
0,8
1
1,2
1,4
1,6
Offered load (normalized)
b)
NSFNET, 32 wavelengths
1,E+00
Burst loss probability
1,E-01
1,E-02
1,E-03
1,E-04
SPR (1LSP)
PER (2LSPs)
1,E-05
PER (4LSPs)
PER (6LSPs)
1,E-06
0,4
0,6
0,8
1
1,2
Offered load (normalized)
c)
EON, 64 wavelengths
Burst loss probability
1,E+00
1,E-01
1,E-02
SPR (1LSP)
PER (2LSPs)
PER (4LSPs)
PER (6LSPs)
PER (8LSPs)
PER (10LSPs)
1,E-03
1,E-04
0,2
0,25
0,3
0,35
0,4
Offered load (normalized)
Figure 10.3: Burst loss probability in PER, a) SIMPLE (32λ), b) NSFNET (32λ),
and c) EON (64λ).
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
a)
108
SIMPLE, 32 wavelengths
1,E+00
Burst loss probability
1,E-01
1,E-02
1,E-03
1,E-04
SPR (1LSP)
BPR (2LSPs)
1,E-05
BPR (4LSPs)
BPR (6LSPs)
1,E-06
0,4
0,6
0,8
1
1,2
1,4
1,6
Offered load (normalized)
b)
NSFNET, 32 wavelengths
1,E+00
Burst loss probability
1,E-01
1,E-02
1,E-03
1,E-04
SPR (1LSP)
BPR (2LSPs)
1,E-05
BPR (4LSPs)
BPR (6LSPs)
1,E-06
0,4
c)
1,2
EON, 64 wavelengths
1,E+00
Burst loss probability
0,6
0,8
1
Offered load (normalized)
1,E-01
1,E-02
SPR (1LSP)
BPR (2LSPs)
BPR (4LSPs)
BPR (6LSPs)
BPR (8LSPs)
BPR (10LSPs)
1,E-03
1,E-04
0,2
0,25
0,3
0,35
0,4
Offered load (normalized)
Figure 10.4: Burst loss probability in BPR, a) SIMPLE (32λ), b) NSFNET (32λ),
and c) EON (64λ).
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
a)
SIMPLE, 32 wavelengths, 0.8 load
1,E+00
Percentage of bursts [x100%]
109
SPR
BPR (2LSPs)
BPR (4LSPs)
BPR (6LSPs)
BPR (8LSPs)
1,E-01
1,E-02
1,E-03
1,E-04
1,E-05
1,E-06
1
b)
3
5
7
9
Number of hops
11
13
15
NSFNET, 32 wavelengths, 0.8 load
Percentage of bursts [x100%]
1,E+00
SPR (1LSP)
BPR (2LSPs)
BPR (4LSPs)
BPR (6LSPs)
BPR (8LSPs)
1,E-01
1,E-02
1,E-03
1,E-04
1,E-05
1
c)
5
7
9
Number of hops
11
13
15
NSFNET, 32 wavelengths, 0.5 load
1,E+00
Percentage of bursts [x100%]
3
SPR
BPR (2LSPs)
BPR (4LSPs)
BPR (6LSPs)
BPR (8LSPs)
1,E-01
1,E-02
1,E-03
1,E-04
1,E-05
1
2
3
4
5
Number of hops
6
7
Figure 10.5: Amount of bursts experiencing given number of hops in BPR, a) SIMPLE
(32λ, ρ = 0.8), b) NSFNET (32λ, ρ = 0.8), and c) NSFNET (32λ, ρ = 0.5).
Chapter 10. Adaptive routing strategies for labelled E-OBS networks
110
node can take a decision according to the state of its own output ports. The route is
selected for each burst individually in all nodes. Although the solution is suboptimal,
since it only considers local information, still it provides good flexibility as well as no
additional signalling is required.
We propose and evaluate two isolated alternative routing algorithms for labelled
E-OBS networks, namely the path excluding routing and the bypass routing. The
obtained results show that our solutions can help in the burst blocking problem in
OBS networks. In particular, BPR can offer a significantly improved performance,
with respect to the shortest-path routing, in small and medium-size networks and
under low and moderate traffic loads. Although the performance of PER is slightly
worsen in such scenarios (comparing to BPR), still, it behaves better under high
loads.
An E-OBS architecture gives a special opportunity to BPR as long as there is
no restriction imposed by the setup of offset times in the edge node on the length of
routing path. Indeed the burst routing path in BPR can be lengthen significantly due
to the deflection operation. Thus the application of this routing strategy in C-OBS
might be difficult.
Alternative routing strategies introduce the problem of out-of-order burst arrival.
Indeed the bursts which are deflected over the paths of different length may arrive
to the destination in an unsettled sequence. The BPR algorithm, which introduce
an unlimited deflection, is particularly sensitive to this problem. Another important
issue is the increase of burst delay; we have already commented that the propagation
delay is a dominant delay factor in OBS. For all these reasons BPR might require
some additional constraints on the maximum number of deflections allowed. As well,
the application of BPR might be reasonable only in low-loaded networks, where the
percentage of deflected bursts is small.
In order to support the PER algorithm in the out-of-order burst arrival problem
we could try to establish the LSPs of similar lengths. In this way the deflected bursts
would experience comparative transmission delays as on the primary paths.
Chapter 11
Optimization of multi-path routing
This Chapter addresses the problem of routing optimization in OBS networks. We
use a simplified analytical model of OBS network with overall burst loss probability
as the primary metric of interest. An approximated form of the overall burst loss
probability, which can be found e.g., in [RVZW03], has a nonlinear character and it
may produce some difficulties in formulating an optimization problem. Indeed the
routing solutions presented in [ZLW+ 04], [HN04], or [TR05] use a linear programming
(LP) formulation, which either does not consider the overall burst loss probability as
a metric of interest or it takes an approximated form of this metric.
In this Chapter we formulate a non-linear optimization problem for multi-path
source routing in OBS and we propose two different methods to solve it. First approach is based on a non-reduced link load calculation with strict partial derivatives
given. The second method is designed for an OBS network model with a reduced link
load calculation. This approach applies a routing optimization framework considered
initially for circuit-switched (CS) networks.
In our routing scenario we assume that there is a pre-established virtual path
topology consisting of a limited number of paths between each pair of source-destination
nodes. We calculate a traffic splitting vector that determines the distribution of traffic over these paths. The proposed solution can be used, in particular, for a static
(pre-planed) routing, where the traffic distribution is calculated based on a given
(long-term) matrix of demands. Then either a periodic or a threshold-triggered update of the splitting vector can be performed if the demand matrix is changed.
11.1
Routing scenario
Consider an OBS network such as that illustrated in Figure 11.1. There are K
links, labelled e = 1, 2, . . . K, and link e comprises Ce wavelengths. A subset p ⊆
{1, 2, . . . K} identifies a path; we define an incidence coefficient αep such that αep = 1
if link e belongs to path p, and αep = 0 otherwise. In the network there is a set P
of paths pre-established between sources (s) and destinations (d). A subset Psd ⊆ P
identifies all paths from node s to node d (later |P | indicates the number of paths in
set P ).
111
Chapter 11. Optimization of multi-path routing
112
path 1
3
2
x1
burst
1
A
4
x2
5
6
path 2
Figure 11.1: Example of OBS network with multi-path source-based routing; x1 and
x2 are the splitting factors and x1 + x2 = 1.
We assume that the routing decision is source-based so that the source node
determines the path of a burst that enters the network (see Figure 11.1). Moreover,
the network applies multi-path routing strategy, i.e., each subset Psd comprises a
small number of paths and a burst can take one of those paths. The path selection
is performed according to a given splitting factor xp , such that the sum of xp of all
the paths p belonging to a given subset Psd is equal to 1.
We assume that the nodes are capable to perform a full wavelength conversion
according to the random wavelength-selection algorithm. A burst going over a path p
is blocked and lost if on a given link k that belongs to p there are no free wavelengths.
Otherwise a wavelength in the link is reserved for the burst duration and then released
immediately after the burst transmission.
The reservation (holding) periods on each link are i.i.d random variables with the
mean equal to the mean burst duration l; for simplicity we assume l = 1. The demand
traffic pattern is described by matrix [tsd ] and bursts destined to a given node d arrive
to a node s as a Poisson process of (long-term) rate tsd /l = tsd . Let tp = tsd for each
p ⊂ Psd . Thus traffic vp offered to path p can be calculated as
vp = xp tp .
(11.1)
Here vector x̄ = (x1 , . . . , x|P | ) determines the distribution of traffic over the network and it should be selected so that to reduce congestion and to improve overall
performance.
11.2
Formulation
11.2.1
Loss models of OBS network
A loss model of OBS network based on the Erlang fixed-point approximation was
proposed by Zukerman in [RVZW03]. In particular, the traffic offered to link e is
Chapter 11. Optimization of multi-path routing
113
obtained as a sum of the traffic offered to all the paths that cross this link diminished
by the traffic lost in the preceding links along these paths,
ρe =
X
αep vp
K
Y
(1 − βpge Eg ) ,
(11.2)
g=1
p∈P
where βpge equals 1 or 0 depending whether or not link g precedes link e along path
p, respectively. We call this model a reduced link load (R-LL) model.
The Zukerman formulation may bring some difficulty in the context of computation of partial derivatives (for optimization purposes). Therefore we propose a
simplified link load model, later called a non-reduced link load (NR-LL) model,
where the traffic offered to link e is calculated as a sum of the traffic offered to all
the paths that cross this link,
X
ρe =
αep vp .
(11.3)
p∈P
The rationale for this proposal lays behind the fact that under low link losses Eg ,
as one can expect in a well dimensioned network, model (11.2) can be approximated
to (11.3).
The main modelling steps include the calculation of burst loss probabilities in
links, paths and entire network, successively.
1. We assume that the offered burst traffic at each link is the aggregation of a large
number of independent traffic flows. Hence, the link range dependence within
the aggregate traffic will be reduced to zero or to very short range dependent
and the traffic arrival at each link in the network can be approximated by the
Poisson process [LLGC06]. Then burst loss probabilities Ee in links are given
by the Erlang loss formula
"C
#−1
e
i
e
X
ρC
ρ
e
Ee = E(ρe , Ce ) = e
.
(11.4)
Ce ! i=0 i!
2. Given the difficulty in obtaining an exact path-level blocking formulas we have
assumed that each blocking event occurs independently from link to link along
any path inside the network. Then loss probabilities Lp of bursts offered to
paths are calculated taking into account the losses in each link that is crossed
by given path, according to the formula
Lp = 1 −
K
Y
(1 − αep Ee ) .
(11.5)
e=1
3. The overall burst loss probability B , which is the sum of traffic lost in the
network normalized by the traffic offered to the network, is obtained as
"
#−1
X
X
.
(11.6)
B=
vp Lp
vp
p∈P
p∈P
Chapter 11. Optimization of multi-path routing
114
a) non-reduced OBS link load model
re=v1+...+vp
v1
vp
...
link e
b) reduced OBS link load model
re=vp(1-Ei)(1-Ej)
vp
Ei
El
Ej
link e
Em
c) reduced CS link load model
re=vp(1-Ei)(1-Ej)(1-El)(1-Em)
vp
Ei
El
Ej
Em
link e
Figure 11.2: Link load models: a) non-reduced OBS, b) reduced OBS, and c) reduced
CS.
Later we will also refer to a reduced link load model of the circuit-switching
network, so here we introduce it. The only difference between this model and the
OBS network ones is the way the link load is calculated. Particularly, the traffic
offered to link e is obtained as a sum of the traffic offered to all the paths that cross
this link diminished by the traffic lost in both the preceding and the succeeding links
along these paths,
ρe =
X
p∈P
K
Y
αep vp
(1 − αep Eg )
(11.7)
g=1,g6=e
= (1 − Ee )−1
X
αep vp (1 − Lp ).
p∈P
The calculation of link, path and overall blocking probabilities is the same as for
Chapter 11. Optimization of multi-path routing
115
NR-LL vs. R-LL model (SP routing)
Burst Loss Probability, B
1,E+00
EON
(64 wavelengths)
1,E-01
NSFNET
(32 wavelengths)
1,E-02
SIMPLE
(8 wavelengths)
1,E-03
R-LL
NR-LL
1,E-04
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
offered load
Figure 11.3: Accuracy of NR-LL model, in SIMPLE, NSFNET, and EON topologies,
with 8, 32, and 64 wavelengths per link, respectively.
the OBS loss model, using formulae (11.4)-(11.6).
Figures 11.2a-c present illustrative examples of link load calculations in all of the
introduced OBS and CS network loss models.
Figure 11.3 compares the overall burst loss probability B of both OBS network
loss models, calculated in the function of traffic load, in different network scenarios
(see Section 10.1 for more details on the evaluation scenario). The network routing
we use here is based on a single-path shortest-hop routing algorithm. We can see that
the accuracy of NR-LL model is very strict for B below 10−2 .
11.2.2
Optimization problem
From equations (11.1) and (11.6) we define a cost function to be the subject of
optimization:
X
B(x̄) =
xp tp Lp
(11.8)
p∈P
The optimization problem is formulated as follows:
min B(x̄)
subject to:
(11.9)
Chapter 11. Optimization of multi-path routing
X
116
xp = 1, ∀Psd ,
(11.10)
p∈Psd
0 ≤ xp ≤ 1, ∀p ∈ P.
(11.11)
Since the overall BLP is a non-linear function of vector x̄ the cost function is
non-linear as well. According to [PW85] for solving such optimization problem we
can use for instance the modified reduced gradient method described in [Har76].
11.3
Partial derivatives
Gradient methods need to employ the calculation of partial derivatives of the cost
function. The partial derivatives indicate a direction which leads to the reduction in
total cost.
11.3.1
NR-LL model
Direct calculation
In NR-LL model the partial derivatives can be derived from formulae (11.3)-(11.6),
successively, starting from
• partial derivatives of the offered traffic load function in link e
∂ρe (x)
= Aei ti =
∂xi
½
ti , if αei = 1
0 , if αei = 0
¾
,
(11.12)
• partial derivatives of the burst loss probability function in link e
µ
¶
∂Ee (x)
Ce − ρe ∂ρe (x)
= Ee Ee +
,
∂xi
ρe
∂xi
(11.13)
• partial derivatives of the burst loss probability function in path p
K
K
Y
∂Lp (x) X
∂Ee (x)
=
αep
(1 − αgp Eg )
,
∂xi
∂xi
e=1
g=1,g6=e
(11.14)
• and finally, partial derivatives of the cost function
X
∂B(x̄)
∂Lp (x)
= ti Li +
tp xp
.
∂xi
∂xi
p∈P
(11.15)
Chapter 11. Optimization of multi-path routing
117
Fast calculation
This direct calculation of partial derivatives may be time-consuming. Indeed we
should find |P | partial derivatives (11.14) to calculate (11.15); notice that |P | may
be really big in larger networks. Instead we provide the following exact derivation,
similar to the one proposed by Kelly in [Kel88] for circuit-switching networks.
We shall write f = f (v; C) when we wish to emphasize the functional dependence
of a function f on the system parameters v = (vp , p ∈ P ) and C = (C1 , . . . , CK ). In
summations, products and the definitions of matrices i and p ranges over P , and e or
g range over {1, 2, . . . , K}.
For each link e define
ηe = E(ρe , Ce − 1) − E(ρe , Ce ).
(11.16)
From (11.4) it follows that
d
E(ρe , Ce ) = [1 − E(ρe , Ce )] ηe .
dρe
(11.17)
d
Ee (v; C) = αei (1 − Ee )ηe .
dvi
(11.18)
From this and (11.3)
Define the form
B(v; E; C) =
X
Ã
vp
p
1−
Y
!
(1 − αgp Eg ) .
(11.19)
g
We find that
Y
∂
B(v; E; C) = 1 −
(1 − αgi Eg ) = Li
∂vi
g
(11.20)
X
∂
αep vp (1 − Lp ).
B(v; E; C) = (1 − Ee )−1
∂Ee
p
(11.21)
and
For each link e define ce such that
X
ce = ηe
αep vp (1 − Lp ).
p
(11.22)
Chapter 11. Optimization of multi-path routing
118
From the above
∂
d
B(v; C) =
B(v; E(v; C); C)
∂vi
dv
" i
#
X
∂
d
∂
=
B(v; E; C)
+
Ee (v; C)
∂vi
dvi
∂Ee
e
X
∂
= Li +
αei (1 − Ee )ηe
B(v; E; C)
∂E
e
e
X
X
αep vp (1 − Lp )
αei ηe
= Li +
e
= Li +
X
(11.23)
p
αei ce .
e
Finally, since (11.1) we have
"
d
B(x̄) = ti Li +
dxi
X
#
αei ce .
(11.24)
e
This calculation of partial derivatives is straightforward. Indeed once K unknowns
(ce ) are pre-calculated then they can be used in (11.24) to obtain the partial derivatives of B.
11.3.2
R-LL model
Computing of partial derivatives in the R-LL model is even more complex than in
the NR-LL model. Therefore in order to find them we take the approach considered
by Kelly in [Kel88] for the circuit-switched network model (see (11.7)) and use it as
a rough approximation.
Let c = (c1 , c2 , . . . , cK ) be the (unique) solution to the equation:


X
X
ce = ηe (1 − Ee )−1
xp tp (1 − Lp ) 1 −
cg 
(11.25)
p:e∈p
g∈p−{e}
Then
"
Ã
d
B(x̄) ≈ ti 1 − (1 − Li ) 1 −
dxi
X
!#
ce
(11.26)
e∈i
Notice that the formula 11.26 corresponds strictly to the CS network case.
11.3.3
Remarks
Although the correctness of our approximation of partial derivatives for R-LL model
is not confirmed theoretically, still our numerical results show that these derivatives
lead us to an optimal solution of the optimization problem. Some explanation of this
Chapter 11. Optimization of multi-path routing
119
fact could be the similarity of both OBS and CS reduced link loss models (see Figure
11.2). Indeed the only difference is that the link load reduction in CS networks is
higher by the traffic lost in succeeding links when comparing to OBS networks.
Regarding the NR-LL model, although we are not able to prove that the (not
unique) solution is optimal in a global sense, numerical results show that several
repetitions of the optimization of (11.8) using formula (11.24) always give us the
same (with a finite numerical precision) near-optimal value of B.
In order to get insight into the character of function B in NR-LL model we calculate it for vector x̄0 (γ), such that:
x̄0 (γ) = γ x̄1 + (1 − γ)x̄2
(11.27)
where x̄1 and x̄2 are two (different) near-optimal vectors, and γ ∈ [0, 1].
Numerical results show that B(x̄0 (γ)) is a monotonic function of near-horizontal
character.
11.4
Implementation issues
The proposed optimization framework can be used to calculate a traffic splitting
vector that determines the distribution of traffic over the network in a multi-path
source based routing scenario. We assume that there is a virtual path topology preestablished that comprise, for instance, a limited number of shortest paths between
each pair of source-destination nodes. Such virtual topology can be established, e.g.,
in a labelled OBS network (see Chapter 9).
Centralized routing
Centralized routing optimization can be applied, for instance, in a pre-planed routing, where the traffic distribution is calculated based on a given (long-term) matrix
of demands. Then either a periodic or a threshold-triggered update of the splitting
vector can be performed if the matrix of demands changes. In principle, both NRLL model and R-LL model can be used for such a centralized routing optimization.
Nevertheless, as long as the accuracy of NR-LL model is very strict at a low burstloss working point and the calculation of its partial derivatives is straightforward this
model is a preferable candidate for centralized routing optimization.
Distributed routing
A distributed routing should react rapidly to a local disturbance at the point of
the disturbance, with slower adjustments in the rest of the network. Similarly like it
was proposed for circuit-switched networks [Kel88] the R-LL model could potentially
be used in a distributed adaptive routing algorithm in OBS networks (some similar
study was presented in [LLGC06]).
In such distributed adaptive routing the network should offer the possibility of
limited communication between the nodes. The nodes should be capable to measure
Chapter 11. Optimization of multi-path routing
120
the loads carried through the links and the source nodes should be able to measure the
loads carried on the paths. Moreover, such routing requires a (limited) arithmetical
processing ability for each link and route, which may be distributed over the nodes of
the network; for example the processing for routes might be carried out at the sources
nodes. Then the measurements of actual loads together with computing of partial
derivatives could be used to implement a decentralized hill-climbing search procedure
able gradually to vary routing patterns in response to changes in the demands placed
on the network (as in [Kel88]).
The design of an optimized distributed routing algorithm is left for future study.
11.5
Performance
We evaluate the performance of our optimized multi-path routing in the simulation
scenario described in Section 10.1. In order to find a splitting vector x̄ that yields
to a near-optimal routing we use a solver fmincon, for constrained nonlinear multivariable functions, which is available in the Matlab environment. Then we apply this
vector in the simulator. The optimized routing, OR-NR and OR-R respectively for
NR-LL model and R-LL model, is compared with simple shortest-hop routing (SPR).
We consider 2 shortest paths per each source-destination pair of nodes; they are not
necessarily disjoint. In SPR only 1 path is available. Uniform traffic matrix as well
as exponential burst inter-arrivals and durations are considered.
In Figure 11.4 we show the overall BLP in the function of offered traffic load ρ
normalized to the link capacity. We evaluate 3 scenarios of small (SIMPLE), medium
(NSFNET) and large (EON) network dimension. We can see that optimized routing
achieves significantly lower losses than SPR in each scenario. Moreover, we can also
observe that both the OR-NR and the OR-R offer the same routing performance.
Finally, we validate that the analytical results (OR-NR (an) in the Figure) calculated
from the model match very well the simulation ones (OR-NR (sim)).
11.5.1
Comparison of routing schemes
Having validated the optimized, with NR-LL model, multi-path routing (OR) we
compare it with PER and BPR isolated alternative routing strategies proposed in
Chapter 10. We consider k = 2 LSPs per each pair of source-destination nodes in
OR, whilst k = 2, or k = 6 in the case of alternative routing.
In Figure 11.5 we evaluate the overall BLP performance in the function of offered,
normalized traffic load. Our first observation is that with the same number of paths
k available and under either low or high load conditions the OR performs better than
the corresponding alternative routing algorithms. The fact can be explained by a
better global knowledge of the network congestion state in the optimized multi-path
routing then in the isolated alternative routing. This knowledge allows to distribute
the traffic over the paths that traverse underutilized links of the network, and so it
preserves from the use of overloaded links.
Chapter 11. Optimization of multi-path routing
a)
121
SIMPLE, 32 wavelengths
1,E-01
Burst loss probability
1,E-02
1,E-03
1,E-04
SPR (sim)
OR-NR (sim)
1,E-05
OR-R (an)
OR-NR (an)
1,E-06
0,5
0,6
0,7
0,8
Offered load (normalized)
b)
NSFNET, 32 wavelengths
1,E-01
Burst loss probability
1,E-02
1,E-03
1,E-04
SPR (sim)
OR-NR (sim)
1,E-05
OR-R (an)
OR-NR (an)
1,E-06
0,4
c)
0,5
0,6
0,7
Offered load (normalized)
0,8
EON, 64 wavelengths
Burst loss probability
1,E-01
1,E-02
1,E-03
SPR (sim)
1,E-04
OR-NR (sim)
OR-NR (an)
1,E-05
0,2
0,3
0,4
Offered load (normalized)
Figure 11.4: Burst loss probability in OR, a) SIMPLE (32λ), b) NSFNET (32λ), and
c) EON (64λ).
Chapter 11. Optimization of multi-path routing
a)
122
SIMPLE, 32 wavelengths
1,E+00
Burst loss probability
1,E-01
1,E-02
1,E-03
SPR (1LSP)
BPR (2LSPs)
BPR (6LSPs)
PER (2LSPs)
PER (6LSPs)
OR (2LSPs)
1,E-04
1,E-05
1,E-06
0,4
0,6
0,8
1
1,2
1,4
1,6
Offered load (normalized)
b)
NSFNET, 32 wavelengths
1,E+00
Burst loss probability
1,E-01
1,E-02
1,E-03
SPR (1LSP)
BPR (2LSPs)
BPR (6LSPs)
PER (2LSPs)
PER (6LSPs)
OR (2LSPs)
1,E-04
1,E-05
1,E-06
0,4
0,6
0,8
1
1,2
Offered load (normalized)
c)
EON, 64 wavelengths
1,E+00
Burst loss probability
1,E-01
1,E-02
1,E-03
SPR (1LSP)
BPR (2LSPs)
BPR (10LSPs)
PER (2LSPs)
PER (10LSPs)
OR (2LSPs)
1,E-04
1,E-05
0,2
0,25
0,3
0,35
0,4
Offered load (normalized)
Figure 11.5: Comparison of optimized multipath source routing with isolated alternative routing strategies, a) SIMPLE (32λ), b) NSFNET (32λ), and c) EON (64λ).
Chapter 11. Optimization of multi-path routing
123
In a small network (Figure 11.5a), both alternative routing algorithms can take
advantage of their reactive contention resolution feature if the number of LSP they can
access is high (k = 6). On the contrary, isolated alternative routing might have some
difficulty with the reduction of burst blocking in larger networks (Figures 11.5b-c).
11.6
Summary
In this Chapter we propose a non-linear optimization method for multi-path source
routing problem in OBS networks. In our proposal we calculate the traffic splitting
vector that determines a near-optimal distribution of traffic over routing paths. The
formulas for partial derivatives we present are straightforward and very fast in computing; it makes the proposed non-linear optimization method a viable alternative
for linear programming formulations.
The simulation results demonstrate that in a static traffic scenario our optimization method effectively distributes the traffic over the network. As a result the
network-wide burst loss probability is reduced compared to the shortest path routing.
Moreover, the optimized multi-path routing outperforms alternative routing strategies if the same number of routing paths is considered.
Our optimization method can be possibly extended to a distributed routing scenario. The design of a distributed adaptive routing algorithm based on the reduced
link load model, and adequate to the one proposed by Kelly for the circuit-switched
networks, is left for future study.
PART V
Conclusions and future works
Chapter 12
Conclusions and future works
The tremendous growth of the Internet, together with the fact that it is a packetbased network, is the main drivers to develop a data-centric transport network. In this
context the optical burst switching architecture is considered as a promising network
solution. The advantage of having small switching granularities is in the conventional
OBS architectures counterbalanced by high burst blocking probabilities. Therefore,
there is a strong requirement for dedicated hardware and control solutions in order to
enable both feasible and effective operation in such networks. In this dissertation we
propose an architectural solution for OBS networks and we addresses several issues
related to QoS provisioning and routing as well as network modelling and performance
evaluation.
E-OBS architecture: characteristics and modelling
We confront the conventional OBS (C-OBS) architecture with the offset timeEmulated OBS (E-OBS) scheme. E-OBS introduces the offset times artificially, by
means of additional fiber delay elements used in core nodes. We show that C-OBS
architectures possess several drawbacks, such as the problem of unfairness in access
to transmission resources, constraints in alternative routing, a need for complex void
filling-based resources reservation algorithms, some difficulties in QoS provisioning,
etc. On the contrary, the E-OBS can bring significant facilities to the mentioned
problems. Taking into account all the arguments provided in this thesis, there is
a motivation for recognizing the E-OBS architecture as an efficient and functional
alternative to conventional OBS networks.
Since E-OBS architectures need for additional fiber delay elements, we provide a
study on their feasibility in relation to other key system parameters. In particular, we
address the problem of congestion in the control plane and the resulting insufficient
offset effect. In order to approach this issue a queuing model of OBS control plane
operation is studied. We give some preliminary results for an exemplary E-OBS
system with a single processor performing at the node controller. Depending on the
distribution of processing times we model such system either as M/M/1 queue with
reneging or as M/D/1/K queue without reneging. The obtained results show that
an appropriate setup of burst lengths may effectively limit the congestion in control
127
Chapter 12. Conclusions and future works
128
plane. Moreover for the analyzed node controller of moderate processing times we
show that the offset times can be provisioned effectively in core nodes, and still, the
performance is preserved.
QoS provisioning
We study the performance of the most addressed mechanisms providing relative
QoS differentiation in OBS networks. In particular, we show that the burst preemptive mechanism concurrently achieves efficient resources utilization and offers highly
effective QoS differentiation. The offset time differentiation mechanism, which is frequently invoked in literature, provides high HP class performance as well, however its
scheduling efficiency, and so the throughput, is aggravated by the variation of offsettimes. Finally, the wavelength threshold-based mechanism, which is however able to
perform class differentiation, is characterized by the poorest overall performance that
significantly depends on its threshold value. The application of this mechanism may
be reasonable only in highly dimensioned networks where the wavelength threshold
is relatively high (in order to service efficiently the LP traffic) and it adapts to traffic
changes.
The high performance of burst preemption mechanism designates it to be a suitable mechanism for QoS differentiation in OBS. Although in this thesis we concern on
relative quality guarantees, still, the preemption scheme can be extended to absolute
QoS provisioning. Such a study can be found e.g. in [OS06]. There the superiority
of a preemptive-based burst dropping mechanism over other mechanisms, also over
the scheme with intentional packet dropping, was demonstrated again. The main
drawback of the burst preemption mechanism in OBS is the overbooking of resources
in case of a successful preemption. Nevertheless, as we have discussed in this thesis,
such a problem can be avoided in E-OBS with the preemption window mechanism
introduced.
Routing
An E-OBS architecture gives a special opportunity to the alternative routing as
long as there is no restriction imposed by the setup of offset times in the edge node
on the length of routing path. As a result, the burst can be freely deflected in
intermediate nodes with any routing algorithm. In this thesis we propose and evaluate
two isolated alternative routing algorithms for labelled E-OBS networks, namely the
path excluding routing and the bypass routing. The obtained results show that our
solutions can help in the burst blocking problem in OBS networks.
Another routing approach that we address in this thesis is multi-path routing. In
this context we propose a novel approach to its optimization. Our proposal is based
on the theory of non-linear optimization with a straightforward calculation of partial
derivatives. Simulation results demonstrate that the optimized routing effectively
reduces the overall burst loss probability with respect to the shortest path routing.
Moreover, if we consider the same number of routing paths available, it outperforms
the alternative routing as well.
Chapter 12. Conclusions and future works
129
As a final remark, we would like to mention that there is one more great benefit
of E-OBS. In particular E-OBS can be seen as an immediate migration step towards
OPS. Indeed due to the application of fiber delay elements the operation of E-OBS
and OPS are very similar. Still, the differences between both technologies lay in
the length of transmission units (burst vs. packet) and signalling mode (out-of-band
vs. in-band), and thus, higher hardware and processing requirements of the OPS.
Nevertheless, as the progress in optical technologies will continue these dissimilarities
should disappear in the future. As a result, the application of E-OBS may facilitate
the migration from OBS networks to OPS networks.
Some particular conclusions of this thesis are the following:
• The problem of unfairness in C-OBS networks starts to play role if the length
of bursts is short, when comparing to the length of offset times. E-OBS is free
of the unfairness problem.
• The burst loss and the delay performance of both C-OBS and E-OBS is (almost)
the same.
• In an exemplary E-OBS node with a single-processor controller, a feasible fiber
delay coil (25µs of delay), and with fast processing times (Tp = 200ns), the
minimum value of average burst length, which is equal to several kbytes, is very
close to the one determined by the control-plane stability constraint. Under
moderate processing times (Tp = 1µs), the length of burst is more restricted,
and it should be 100kbytes at least (instead of 77kbytes obtained from the
stability constraint).
• Effective throughput of the burst preemption mechanism is higher than of the
offset time differentiation mechanism. Under high loads (ρ = 0.8) the difference
might be even of 2 ÷ 3%.
• The amount of (useless) phantom bursts generated in a single node, which is
enhanced with a burst preemption mechanism, is of about 4% in an exemplary
system of 16 wavelengths, ρ = 0.8, HP load of 30%.
• Application of the preemption window mechanism in an E-OBS node allows
to avoid the problem of phantom bursts. In an exemplary OBS system of 32
wavelengths, ρ = 0.8, α = 30%, and the preemption window equal to 15µs
(3km), the HP burst loss probabilities of about 10−5 can be achieved.
• When using isolated alternative routing strategies in highly loaded, small networks, the improvement of burst loss performance can be even of 2 orders of
magnitude, comparing to shortest path routing. In larger networks, the improvement is not so high (below 1 order of magnitude).
• Under the same number of paths available, optimized multi-path routing can
perform better than isolated alternative routing algorithms.
Chapter 12. Conclusions and future works
130
Concluding, E-OBS was shown to be a functional and feasible alternative for OBS
networks, with a support for highly effective QoS provisioning and facilitated routing
management.
We can distinguish several issues that would be addressed in future work.
• Firstly, the modelling of multi-processor switch controller architectures, which is
desired in the context of system design and dimensioning. Such study would allow to find the trade-offs between performance and complexity/cost of different
controller architectures.
• Another topic started in this thesis is optimization of routing in OBS networks.
In this context, both distributed multi-path routing and single-path routing
strategies, possibly with QoS constraints, will be studied. Since multi-path
routing introduces the problem of out-of-order burst arrival this issue has to be
addressed as well.
• An important issue which is a hot topic of current research activity is the
deployment of control plane in OBS networks. As a solution, we have initiated
to consider the generalized MPLS protocol (GMPLS). Adaptation of GMPLS to
OBS might be desired, in particular, in the context of the network migration as
long as GMPLS is an accepted solution in OCS networks. While GMPLS should
facilitate the coexistence of OCS and OBS, the concept of E-OBS enables the
migration towards OPS networks. Therefore, the loop can be closed allowing
the continuous deployment of ASON, OBS and OPS.
Acronyms
ABT
ADSL
ASON
ATM
BCP
BD-W
BLP
BP
BPR
CC
C-OBS
CP
CPU
CS
DWDM
E-OBS
FDC
FDL
FTTH
GMPLS
HP
IP
LP
LSP
MEMS
MPLS
NG-SDH
NLP
NR-LL
OBS
OCS
ODM
OPS
OR
ORION
ATM Block Transfer
Asymmetric Digital Subscriber Line
Automatic Switched Optical Network
Asynchronous Transfer Mode
Burst CP
Burst Dropping with Wavelength threshold
Burst Loss Probability
Burst Preemption
Baypass Path Routing
Control Channel
Conventional OBS
Control Packet
Control Processor Unit
Circuit Switching
Dense WDM
Offset Time Emulated OBS
Fiber Delay Coil
Fiber Delay Line
Fiber to the Home
Generalized MPLS
High Priority
Internet Protocol
Low Priority Class
Label Switched Path
Micro-Electro-Mechanical Systems
Multi-Protocol Label Switching
Next-Generation SDH
Non-Linear Programming
Non-Reduced Link Load
Optical Burst Switching
Optical Circuit Switching
Optical Drop Multiplexer
Optical Packet Switching
Optimized Routing
Ontario Research and Innovation Optical Network
131
Acronyms
OR-NR
OR-R
OT
OTD
OXC
P2P
PER
PW
QoS
RAM
RED
R-LL
RWA
SDH
SOA
SONET
SP
SPR
TAG
TAW
TCP
TE
UDP
WDM
WLAN
WR-OBS
WS
132
OR with NR-LL model
OR with R-LL model
Offset Time
Offset Time Differentiation
Optical Cross-connect
Pear to Pear
Path Excluding Routing
Preemption Window
Quality of Service
Random Access Memory
Random Early Detection
Reduced Link Load
Routing and Wavelength Assignment
Synchronous Digital Hierarchy
Semiconductor Optical Amplifier
Synchronous Optical Networking
Shortest Path
SP Routing
Tell-and-Go
Tell-and-Wait
Transmission Control Protocol
Traffic Engineering
User Datagram Protocol
Wavelength Division Multiplexing
Wireless Local Area Network
Wavelength-Routed OBS
Wavelength Conversion/Wavelength Converter
Bibliography
[ACP04]
A. Agusti and C. Cervello-Pastor. A new contentionless dynamic
routing protocol for obs using wavelength occupation. In Proceedings of the 12th IEEE Mediterranean Electrotechnical Conference
(MELECON 2004), Dubrovnik, Croatia, May 2004.
[AdDA07]
C. G. Argos, O. Gonzlez de Dios, and J. Aracil. Adaptive multi-path
routing for obs networks. In Proceedings of the 9th IEEE International Conference on Transparent Optical Networks (ICTON2007),
Rome, Italy, July 2007.
[AEBS05]
A. Agrawal, T. S. El-Bawab, and L. B. Sofman. Comparative account of bandwidth efficiency in optical burst switching and optical circuit switching networks. Photonic Network Communications,
9(3):297309, 2005.
[AST+ 06]
A. Al Amin, K. Shimizu, M. Takenaka, T. Tanemura, R. Inohara,
and K. Nishimura et al. 40/10 gbps bit-rate transparent burst switching and contention resolving wavelength conversion in an optical
router prototype. In Proceedings of 32nd European Conference on
Optical Communications (ECOC 2006), Cannes, France, October
2006.
[Bar57]
D. Y. Barrer. Queuing with impatient customers and ordered service.
Journal Operations Res. Soc. Amer., 5:650–656, 1957.
[BBE+ 05]
I. Baldine, A. Bragg, G. Evans, M. Pratt, M. Singhai, D. Stevenson, and R. Uppalli.
Jumpstart deployments in ultra-highperformance optical networking testbeds. IEEE Optical Communications, 43(11):S18–S25, November 2005.
[BD06]
N. Barakat and T. E. Darcie. Control-plane congestion in obs networks. In Proceedings of the 4th Workshop on OBS (WOBS), San
Jose, CA (USA), October 2006.
[BD07]
N. Barakat and T. E. Darcie. The control-plane stability constraint
in optical burst switching networks. IEEE Communications Letters,
11(3):267–269, March 2007.
133
Bibliography
134
[BIPe99]
D. Blumenthal, T. Ikegami, P. R. Prucnal, and (editors). Special issue on photonic packet switching technologies, techniques, and systems. IEEE Journal of Lightwave Technology, 17(12), December
1999.
[BNH+ 03]
S. Bjornstad, M. Nord, D.R. Hjelme, N. Stol, F. Callegati, W. Cerroni, C. Raffaelli, P. Zaffoni, C. M. Gauger, C. Develder, J. Cheyns,
E. Van Breusegem, E. Baert, D. Colle, M. Pickavet, P. Demeester,
M. Lackovic, D. Careglio, G. Junyent, M. Klinkowski, M. Marciniak,
and M. Kowalewski. Optical burst and packet switching: Node and
network design, contention resolution and quality of service. In Proceedings of the 7th IEEE International Conference on Telecommunications (ConTEL2003), Zagreb, Croatia, June 2003.
[Boc05]
S. Bocquet. Queueing theory with reneging. Technical Report of
DSTO, September 2005.
[Bou03]
N. Boudriga. Optical burst switching protocols for supporting
qos and adaptive routing. Computer Communication Journal,
26(15):1804–1812, 2003.
[BP03]
T. Battestilli and H. Perros. An introduction to optical burst switching. IEEE Optical Communications, 41(8):S10–S15, August 2003.
[BS04]
N. Barakat and E.H. Sargent. On optimal ingress treatment of delaysensitive traffic in multi-class obs systems. In Proceedings of the 3rd
International Workshop on Optical Burst Switching (WOBS 2004),
co-located with BroadNets 2004, San Jose, CA (USA), October 2004.
[Cav00]
D. Cavendish. Evolution of optical transport technologies: from
sonet/sdh to wdm. IEEE Communications Magazine, 38:164–172,
June 2000.
[CC01]
F. Callegati and W. Cerroni. Wavelength allocation algorithms in
optical buffers. In Proceedings of IEEE International Conference on
Communications (ICC 2001), Helsinki, Finland, June 2001.
[CCRZ04]
F. Callegati, W. Cerroni, C. Raffaelli, and P. Zaffoni. Wavelength
and time domain exploitation for qos management in optical packet
switches. Computer Networks, 44(1):569–582, January 2004.
[CCXV99]
F. Callegati, H. Cankaya, Y. Xiong, and M. Vandenhoute. Design
issues of optical ip routers for internet backbone applications. IEEE
Communications Magazine, 37(12):124–128, December 1999.
[CDB+ 03]
J. Cheyns, C. Develder, E. Van Breusegem, A. Ackaert, M. Pickavet,
and P. Demeester. Routing in an awg-based optical packet switch.
Photonic Network Communications, 5(1):69–80, 2003.
Bibliography
135
[CEJ05a]
T. Coutelen, H. Elbiaze, and B. Jaumard. An efficient adaptive offset
mechanism to reduce burst losses in obs networks. In Proceedings of
IEEE Global Communications Conference (GLOBECOM 2005), St.
Louis, MO (USA), December 2005.
[CEJ05b]
T. Coutelen, H. Elbiaze, and B. Jaumard. An efficient adaptive offset
mechanism to reduce burst losses in obs networks. In Proceedings of
Proceedings of IEEE Global Communications Conference (GLOBECOM 2005), St. Louis, MO (USA), December 2005.
[CEJ05c]
T. Coutelen, H. Elbiaze, and B. Jaumard. Performance comparison of ocs and obs switching paradigms. In Proceedings of the
7th IEEE International Conference on Transparent Optical Networks
(ICTON2005), , Barcelona, Spain, July 2005.
[CGK92]
I. Chlamtac, A. Ganz, and G. Karmi. Lightpath communications:
An approach to high-bandwidth optical wans. IEEE Transactions
on Communications, 40(7):1171–1182, July 1992.
[Chi95]
D. Chiaroni. A novel photonic architecture for high capacity atm
switch applications. In Proceedings of the Photonics in Switching
conference, Salt Lake City, UT, April 1995.
[Chi01]
D. Chiaroni. Status and applications of optical packet switching.
In Proceedings of 27th European Conference on Optical Communications (ECOC 2001), Amsterdam, Netherlands, October 2001.
[CLP02]
P.B Chu, S.-S. Lee, and S. Park. Mems: The path to large optical crossconnects. IEEE Communications Magazine, 40(3):80–87,
March 2002.
[CMC06]
Q. Chen, G. Mohan, and K.C. Chua. Route optimization for efficient
failure recovery in optical burst switched networks. In Proceedings
of IEEE High Performance Switching and Routing workshop (HPSR
2006), Poznan, Poland, June 2006.
[CR06]
M. Casoni and C. Raffaelli. Analytical framework for end-to-end
design of optical burst-switched networks. Optical switching and
Networking, (4):33–43, August 2006.
[CVG+ 98]
A. Carena, M. D. Vaughn, R. Gaudino, M. Shell, and D. J. Blumenthal. Opera: An optical packet experimental router architecture with
label swapping capability. IEEE Journal of Lightwave Technology,
16(12), December 1998.
[CWXQ03]
Y. Chen, H. Wu, D. Xu, and C. Qiao. Performance analysis of optical
burst switched node with deflection routing. In Proceedings of IEEE
International Conference on Communication (ICC 2003), Seattle,
USA, May 2003.
Bibliography
136
[CZZ04]
C. Cameron, A. Zalesky, and M. Zukerman. Shortest path prioritized random deflection routing (sp-prdr) in optical burst switched
networks. In Proceedings of ICST International Workshop on Optical
Burst Switching (WOBS), San Jose, USA, 2004.
[DG01]
K. Dolzer and C.M. Gauger. On burst assembly in optical burst
switching networks - a performance evaluation of just-enough-time.
In Proceedings of the 17th International Teletraffic Congress (ITC
17), Salvador, Brazil, December 2001.
[DHL+ 03]
H. J. S. Dorren, M. T. Hill, Y. Liu, N. Calabretta, A. Srivatsa, F. M.
Huijskens, H. de Waardt, and G. D. Khoe. Optical packet switching
and buffering by using all-optical signal processing methods. Journal
of Lightwave Technology, 21(1):2–12, January 2003.
[DKKB00]
M. Duser, E. Kozlovski, R.I. Killey, and P. Bayvel. Design tradeoffs in optical burst switched networks with dynamic wavelength
allocation. In Proceedings of 26th European Conference on Optical
Communications (ECOC 2000), Munich, Germany, September 2000.
[DPZQ06]
Y. Du, T. Pu, H. Zhang, and Y. Quo. Adaptive load balancing routing algorithm for optical burst-switching networks. In Proceedings
of Optical Fiber Communication Conference (OFC 2006), Anaheim,
CL (USA), March 2006.
[ELP03]
V. Eramo, M. Listanti, and P. Pacifici. A comparison study on the
number of wavelength converters needed in synchronous and asynchronous all-optical switching architectures. Journal of Lightwave
Technology, 21(2):340–355, February 2003.
[Fib07a]
Compact time delay coils, 2007. http://www.newport.com/.
[Fib07b]
Fiber delay coils, 2007. http://www.generalphotonics.com/.
[FJ03a]
F. Farahmand and J. Jue. Look-ahead window contention resolution
in optical burst switched networks. In Proceedings of IEEE High
Performance Switching and Routing workshop (HPSR 2003), Torino,
Italy, June 2003.
[FJ03b]
F. Farahmand and J. P. Jue. Supporting qos with look-ahead window contention resolution in optical burst switched network. In Proceedings of IEEE Global Communications Conference (GLOBECOM
2003), San Francisco, CA (USA), December 2003.
[G.700]
ITU-T Recommendation G.707. Network node interface for the synchronous digital hierarchy (sdh), October 2000.
Bibliography
137
[Gau02]
C. Gauger. Dimensioning of fdl buffers for optical burst switching
nodes. In Proceedings of the 6th IFIP Working Conference on Optical Network Design and Modelling (ONDM 2002), Torino, Italy,
February 2002.
[Gau03]
C. Gauger. Trends in optical burst switching. In Proceedings of
the SPIE/ITCOM 2003, volume 5247, pages 115–125, Orlando, FL
(USA), September 2003.
[GBIQ04]
S. Ganguly, S. Bhatnagar, R. Izmailov, and C. Qiao. Multi-path
adaptive optical burst forwarding. In Proceedings of IEEE High Performance Switching and Routing workshop (HPSR 2004), Phoenix,
AR (USA), April 2004.
[GDW03]
N. Ghani, S. Dixit, and T.-S. Wang. On ip-wdm integration: a retrospective. IEEE Communications Magazine, 41(9):42–45, September
2003.
[GKS04]
C. Gauger, M. Kohn, and J. Scharf. Performance of contention resolution strategies in obs network scenarios. In Proceedings of the
3rd International Conference on the Optical Internet (COIN2004),
Yokohama, Japan, July 2004.
[GKZM05]
C. Gauger, M. Kohn, J. Zhang, and B. Mukherjee. Network performance of optical burst/packet switching: The impact of dimensioning, routing and contention resolution. In ITG-Fachtagung Photonic
Networks, Leipzig, Germany, May 2005.
[Gun07]
S. Gunreben. Multi-layer analysis to quantify the impact of optical burst reordering on tcp performance. In Proceedings of the
9th IEEE International Conference on Transparent Optical Networks
(ICTON2007), Rome, Italy, July 2007.
[GWL+ 05]
H. Guo, J. Wu, X. Liu, J. Lin, and Y. Ji. Multi-qos traffic transmission experiments on obs network testbed. In Proceedings of 31nd European Conference on Optical Communications (ECOC 2005), Glasgow, Scotland, September 2005.
[GZ06]
D. Gao and H. Zhang. Information sharing based optimal routing
for optical burst switching (obs) network. In Proceedings of Optical
Fiber Communication Conference (OFC 2006), Anaheim, CL (USA),
March 2006.
[HAM+ 05]
Pan H., T. Abe, Y. Mori, Ch. Young-Bok, and H. Okada. Feedbackbased load balancing routing for optical burst switching networks. In
Proceedings of the 11th Asia-Pacific Conference on Communications
(APCC 2005), Perth, Western Australia, October 2005.
Bibliography
138
[Har76]
R.J Harris. The modified reduced gradient method for optimally
dimensioning telephone networks. Australian Telecom. Research,
10(1):30–35, 1976.
[HCA98]
D. K. Hunter, M. C. Chia, and I. Andonovic. Buffering in optical
packet switches. Journal of Lightwave Technology, 16(12):2081–2094,
December 1998.
[HHM05]
Y. Huang, J.P. Heritage, and B. Mukherjee. Dynamic routing
with preplanned congestion avoidance for survivable optical burstswitched (obs) networks. In Proceedings of Optical Fiber Communication Conference (OFC 2005), Anaheim, CL (USA), March 2005.
[HLH02a]
C. Hsu, T. Liu, and N. Huang. Performance analysis of deflection
routing in optical burst-switched networks. In Proceedings of the 21st
Joint Conference of IEEE Computer and Communications Societies
(INFOCOM 2002), New York, NY (USA), June 2002.
[HLH02b]
C. F. Hsu, T. L. Liu, and N. F. Huang. Performance analysis of
deflection routing in optical burst-switched networks. In Proceedings of the 21st Conference of the IEEE Communications Society
(INFOCOM 2002), New York, NY (USA), June 2002.
[HN04]
E. Hyytia and L. Nieminen. Linear program formulation for routing
problem in obs networks. In Proceedings of the 9th IEEE Symposium on Computers and Communications (ISCC 2004), Alexandria,
Egypt, June - July 2004.
[HSE04]
H.Mouftah, S.Said, and H. Elbiaze. A qos-based restoration mechanism for obs networks. In Proceedings of the 8th IEEE International
Conference on Transparent Optical Networks (ICTON 2006), volume 3, pages 1853–1863, Hong Kong, China, March 2004.
[HTM06]
Y. Hirota, H. Tode, and K. Murakami. Cooperation method considering wavelength assignment and routing problem in optical burst
switched networks. In Proceedings of Optical Fiber Communication
Conference (OFC 2006), Anaheim, CL (USA), March 2006.
[I.300]
ITU-T Recommendation I.371. Traffic control and congestion control
in b-isdn, March 2000.
[IA01]
M. Izal and J. Aracil. On the influence of self-similarity on optical
burst switching traffic. In Proceedings of IEEE Global Communications Conference (GLOBECOM 2001), San Antonio, TX (USA),
November 2001.
[IYS05]
D. Ishii, N. Yamanaka, and I. Sasase. Self-learning route selection
scheme using multipath searching packets in an obs network. Journal
of Optical Networking, 4(7):432–445, 2005.
Bibliography
139
[JECA03]
J.Cheyns, E.V.Breusegem, C.Develder, and A.Ackaert. Performance
improvement of an internally-blocking optical/packet switch. In
Proceedings of IEEE International Conference on Communications
(ICC 2003), Anchorage, AK (USA), May 2003.
[JQX00]
M. Jeong, C. Qiao, and Y. Xiong. Reliable wdm multicast in optical
burst-switched networks. In Proceedings of Optical Networking and
Communications Conference (Opticomm 2000), Dallas, TX (USA),
October 2000.
[JXC+ 00]
M. Jeong, Y. Xiong, H. C. Cankaya, M. Vandenhoute, and C. Qiao.
Efficient multicast schemes for optical burst-switched wdm networks.
In Proceedings of IEEE International Conference on Communication
(ICC 2000), New Orleans, LA (USA), June 2000.
[KA98]
E. Karasan and E. Ayanoglu. Performance of wdm transport
networks. IEEE Journal on Selected Areas in Communications,
16(9):1081–1096, September 1998.
[KA03]
A. Kaheel and H. Alnuweiri. A strict priority scheme for quality-of
service provisioning in optical burst switching networks. In Proceedings of IEEE Symposium on Computers and Communications (ISCC
2003), Turkey, June 2003.
[KA04]
A. Kaheel and H. Alnuweiri. Quantitative qos guarantees in labeled
optical burst switching networks. In Proceedings of the IEEE Global
Communications Conference (GLOBECOM 2004), Dallas, TX, November 2004.
[Kar02]
S. V. Kartalopoulos. DWDM: Networks, Devices, and Technology.
John Wiley & Sons, October 2002.
[KB02]
E. Kozlovski and P. Bayvel. Qos performance of wr-obs network architecture with request scheduling. In Proceedings of IFIP 6th Working Conference on Optical Networks Design and Modelling (ONDM
2002), Turin, Italy, February 2002.
[KCK04]
J. Kim, J. Choi, and M. Kang. Offset-time based scheduling algorithm for burst control packet in optical burst switching networks.
In Proceedings of (ICOIN), Busan, Korea, February 2004.
[KCM04]
B. C. Kim, Y. Z. Cho, and D. Montgomery. An efficient optical burst
switching technique for multi-hop networks. IEICE Transactions on
Communications, E87-B(6):1737–1740, June 2004.
[KCMSP06]
M. Klinkowski, D. Careglio, D. Morato, and J. Sole-Pareta. Effective
burst preemption in obs network. In Proceedings of IEEE High Performance Switching and Routing workshop (HPSR 2006), Poznan,
Poland, June 2006.
Bibliography
140
[KCSSP05]
M. Klinkowski, D. Careglio, S. Spadaro, and J. Solé-Pareta. Impact
of burst length differentiation on qos performance in obs networks. In
Proceedings of the 7th IEEE International Conference on Transparent Optical Networks (ICTON 2005), Barcelona, Spain, July 2005.
[Kel88]
F. P. Kelly. Routing in circuit-switched networks: Optimization,
shadow prices and decentralization. Advanced Applied Probability,
20:112–144, 1988.
[KG03a]
M. Khn and C. Gauger. Contention-based limited deflection routing
in obs networks. In Proceedings of Global Telecommunications Conference (GLOBECOM 2003), San Francisco, CL (USA), December
2003.
[KG03b]
M. Khn and C. Gauger. Dimensioning of sdh/wdm multilayer networks. In Proceedings of the 4th ITG Workshop on Photonic Networks, Leipzig, Germany, 2003.
[KHCSP05]
M. Klinkowski, F. Herrero, D. Careglio, and J. Sol-Pareta. Adaptive
routing algorithms for optical packet switching networks. In Proceedings of the 9th IFIP Working Conference on Optical Networks
Design and Modelling (ONDM 2005), Milan, Italy, February 2005.
[KKK02]
S. Kim, N. Kim, and M. Kang. Contention resolution for optical
burst switching networks using alternative routing. In Proceedings
of IEEE International Conference on Communications (ICC 2002),
New York, NY (USA), April-May 2002.
[KM01]
M. Klinkowski and M. Marciniak. Development of ip/wdm optical
networks. In Proceedings of 3th International Workshop on Laser
and Fiber optic Network Modeling (LFNM 2001), Kharkov, Ukraine,
May 2001.
[KMM+ 05]
K. Kitayama, T. Miki, T. Morioka, H. Tsushima, M. Koga, K. Mori,
S. Araki, K. Sato, H. Onaka, S. Namiki, and T. Aoyama. Photonic network r&d activities in japan-current activities and future
perspectives. IEEE Journal of Lightwave Technology, 23(10):2404–
3418, October 2005.
[LES00]
M. Listanti, V. Eramo, and R. Sabella. Architectural and technological issues for future optical internet networks. IEEE Communications
Magazine, 38(9):82–92, September 2000.
[LKSG03]
S. Lee, H. Kim, J. Song, and D. Griffith. A study on deflection
routing in optical burst-switched networks. Photonic Network Communications, 6(1):51–59, 2003.
Bibliography
141
[LLGC06]
J. Lu, Y. Liu, M. Gurusamy, and K.C. Chua. Gradient projection
based multi-path traffic routing in optical burst switching networks.
In Proceedings of IEEE High Performance Switching and Routing
workshop (HPSR 2006), Poznan, Poland, June 2006.
[LMC05]
J. Li, G. Mohan, and Kee Chaing Chua. Dynamic load balancing in
ip-over-wdm optical burst switching networks. Computer Networks,
47(3):393–408, 2005.
[LQXX04]
J. Li, C. Qiao, J. Xu, and D. Xu. Maximizing throughput for optical
burst switching networks. In Proceedings of the 23rd Conference
of the IEEE Communications Society (INFOCOM 2004), volume 3,
pages 1853–1863, Hong Kong, China, March 2004.
[LQYG06]
Xin Liu, Chunming Qiao, Xiang Yu, and Weibo Gong. A fair packetlevel performance comparison of obs and ocs. In Proceedings of Optical Fiber Communication (OFC 2006), Anaheim, CL (USA), March
2006.
[LTWW94]
W. E. Leland, M. S. Taqqu, W. Willinger, and D. V. Wilson. On
the self-similar nature of ethernet traffic. IEEE/ACM Transactions
on Networking, 2(1):1–15, 1994.
[LY06]
J. Li and K. L. Yeung. Burst cloning with load balancing. In Proceedings of Optical Fiber Communication Conference (OFC 2006),
Anaheim, CL (USA), March 2006.
[LYH+ 06]
K. Long, X. Yang, S. Huang, Q. Chen, and R. Wang. Adaptive
parameter deflection routing to resolve contentions in obs networks.
In Proceedings of the 5th International Conference on Networking
(Networking 2006), Coimbra, Portugal, May 2006.
[MGK+ 04]
I. De Miguel, J. C. Gonzalez, T. Koonen, R. Duran, P. Fernandez,
and I. T. Monroy. Polymorphic architectures for optical networks and
their seamless evolution towards next generation networks. Photonic
Network Communications, 8(2):177–189, 2004.
[MRZ04]
G. Muretto, C. Raffaelli, and P. Zaffoni. Effective implementation
of void filling in obs networks with service differentiation. In Proceedings of the 3rd International Workshop on Optical Burst Switching (co-located with BroadNets 2004), San Jose, CA (USA), October
2004.
[Nor03]
M. Nord. Node design in optical packet and optical burst switching. In Proceedings of 5th International Conference on Transparent
Optical Networks (ICTON 2003), Warsaw, Poland, June-July 2003.
Bibliography
142
[Nsf]
Nsfnet-the
national
http://moat.nlanr.net/.
science
foundation
network.
[NTL+ 05]
K. Nashimoto, N. Tanaka, M LaBuda, D. Ritums, J. Dawley, M. Raj,
D. Kudzuma, and T. Vo. High-speed plzt optical switches for
burst and packet switching. In Proceedings of the 2nd International
Conference on Broadband Networks (Broadnets 2005), Boston, MA
(USA), October 2005.
[OA05]
N. Ogino and N. Arahata. A decentralized optical bursts routing based on adaptive load splitting into pre-calculated multiple
paths. IEICE Transactions on Communications, E88-B(12):4507–
4516, 2005.
[Odl04]
A. Odlyzko. Pricing and architecture of the internet: Historical perspectives from telecommunications and transportation. In Proceedings of the 32nd Research Conference on Communication, Information and Internet Policy (TPRC 2004), Arlington, Virginia, October
2004.
[OS06]
Harald Overby and Norvald Stol. Qos differentiation in asynchronous bufferless optical packet switched networks. Wireless Networks,
12(3):383–394, June 2006.
[OTYC05]
L. Ou, X. Tan, H. Yao, and W. Cheng. A study on dynamic load
balanced routing techniques in time-slotted optical burst switched
networks. In Proceedings of the 3rd International Conference on
Networking and Mobile Computing, (ICCNMC 2005), Zhangjiajie,
China, August 2005.
[PMP07]
J. M. Pedro, P. Monteiro, and J. J. O. Pires. Efficient multi-path
routing for optical burst-switched networks. In Proceedings of Conference on Telecommunications (ConfTele), Peniche , Portugal, May
2007.
[PPP03]
G. I. Papadimitriou, Ch. Papazoglou, and A. S. Pomportsis. Optical switching: Switch fabrics, techniques, and architectures. IEEE
Journal of Lightwave Technology, 21(2):384–405, February 2003.
[PPP06]
G. I. Papadimitriou, Ch. Papazoglou, and A. S. Pomportsis. Optical
Switching. Wiley-Interscience, December 2006.
[Pro05]
FP6-506760 IP Nobel Project. Deliverable d16 - preliminary definition of burst/packet network and node architectures and solutions,
March 2005.
[PSPCK07]
P. Pedroso, J. Sol-Pareta, D. Careglio, and M. Klinkowski. Integrating gmpls in the obs networks control plane. In Proceedings of the
Bibliography
143
9th IEEE International Conference on Transparent Optical Networks
(ICTON2007), Rome, Italy, July 2007.
[PW85]
M. Pioro and B. Wallstrom. Multihour optimization of nonhierarchical circuit switched communication networks with sequential routing. In Proceedings of the 11th International Teletraffic
Congress (ITC-11), 1985.
[Qia00]
C. Qiao. Labeled optical burst switching for ip-over-wdm integration.
IEEE Communications Magazine, 38(9):104–114, September 2000.
[QY99]
C. Qiao and M. Yoo. Optical burst switching (obs) - a new paradigm
for an optical internet. Journal of High Speed Networks, 8(1):69–84,
March 1999.
[RFL05]
J. Rodrigues, M.M. Freire Freire, and P. Lorenz. Performance implications of meshing degree for optical burst switched networks using
one-way resource reservation protocols. Telecommunication Systems,
30(1-3):35–47, November 2005.
[RG04]
M. De Vega Rodrigo and J. Gotz. An analytical study of optical
burst switching aggregation strategies. In Proceedings of IEEE/SPIE
Third International Workshop on Optical Burst Switching (WOBS
2004), San Jose, CA (USA), October 2004.
[RI03]
B. Mikac R. Inkret, A. Kuchar. Advanced infrastructure for photonic
networks. extended final report of cost action 266. Technical Report
ISBN 953-184-064-4, Published by the Faculty of Electical Engineering and Computing, University of Zagreb, Croatia, September 2003.
Available also at http://www.ufe.cz/dpt240/cost266/index.html under item ’Reports’.
[Ros01]
E. Rosen. Multiprotocol label switching architecture. IETF RFC
3031, January 2001.
[RVZW03]
Z. Rosberg, H. L. Vu, M. Zukerman, and J. White. Blocking probabilities of optical burst switching networks based on reduced load
fixed point approximations. In Proceedings of IEEE INFOCOM
2003, New York, NY (USA), March-April 2003.
[SC05]
J. M. Smith and F. Cruz. The buffer allocation problem for general
finite buffer queueing networks. IIE Transactions, 37(4):343–366,
September 2005.
[SPG05]
M. Schlosser, E. Patzak, and P. Gelpke. Impact of deflection routing
on tcp performance in optical burst switching networks. In Proceedings of the 7th IEEE International Conference on Transparent
Optical Networks (ICTON2005), Barcelona, Spain, July 2005.
Bibliography
144
[Tan88]
A. S. Tanenbaum. Computer Networks (2nd ed.). Prentice Hall,
1988.
[TR05]
J. Teng and G. N. Rouskas. Traffic engineering approach to path
selection in optical burst switching networks. Journal of Optical
Networking, 4(11):759–777, 2005.
[Tur99]
J.S. Turner. Terabit burst switching. Journal of High Speed Networks, 8(1):3–16, March 1999.
[TVJ03]
G.P. Thodime, V.M. Vokkarane, and J.P. Jue. Dynamic congestionbased load balanced routing in optical burst-switched networks. In
Proceedings of IEEE Global Communications Conference (GLOBECOM 2003), San Francisco, CA (USA), November 2003.
[TZ99]
R. S. Tucker and W. De Zhong. Photonic packet switching: An
overview. IEICE Trans. commun., E82-B(2):254–264, February
1999.
[VGPMGH+ 07] J. Veiga-Gontan, P. Pavon-Marino, J. Garcia-Haro, M. Rodelgo,
C. Lopez-Bravo, and F.J. Gonzalez-Castano. Network processors:
a practical approach for achieving wire-speed packet processing in
the emerging optical backbone networks. In Proceedings of the 9th
IEEE International Conference on Transparent Optical Networks
(ICTON2007), Rome, Italy, July 2007.
[VJ02a]
V. Vokkarane and J. Jue. Burst segmentation: an approach for reducing packet loss in optical burst switched networks. In Proceedings
of IEEE International Conference on Communications (ICC 2002),
New York, NY (USA), April-May 2002.
[VJ02b]
V. Vokkarane and J. Jue. Prioritized routing and burst segmentation for qos in optical burst-switched networks. In Proceedings
of Optical Fiber Communication Conference (OFC 2002), Anaheim,
CA (USA), March 2002.
[VJ03]
V. M. Vokkarane and J. P. Jue. Prioritized burst segmentation and
composite burst-assembly techniques for qos support in optical burstswitched networks. IEEE Journal on Selected Areas in Communications (JSAC), 21(7):1198–1209, September 2003.
[VKM+ 01]
M. Veeraraghavan, R. Karry, T. Moors, M. Karol, and R. Grobler.
Architectures and protocols that enable new applications on optical
networks. IEEE Communications Magazine, 39(3):118–127, March
2001.
[VS97]
E. Varvarigos and V. Sharma. The ready-to-go virtual-circuit protocol: A loss-free protocol for multigigabit networks using fifo buffers.
IEEE/ACM Transactions on Networking, 5(5):705–718, 1997.
Bibliography
145
[Whi83]
W. Whitt. The queueing network analyzer. Bell Systems Technical
Journal, 62(9):2779–2815, November 1983.
[Wid95]
I. Widjaja. Performance analysis of burst admission-control protocols. IEEE Proceeding of Communications, 142:7–14, 1995.
[WM00]
J. Y. Wei and R. I. McFarland. Just-in-time signaling for wdm optical
burst switching networks. IEEE Journal of Lightwave Technology,
18(12):2019–2037, December 2000.
[WMA02]
X. Wang, H. Morikawa, and T. Aoyama. Burst optical deflection
routing protocol for wavelength routing wdm networks. Optical Networks Magazine, 3(6), November/December 2002.
[WR04]
Y. Wang and B. Ramamurthy. Cpq: A control packet queuing optical
burst switching protocol for supporting qos. In Proceedings of 3rd
International Workshop on Optical Burst Switching (WOBS 2004),
co-located with BroadNets 2004, San Jose, CA (USA), October 2004.
[WZSZ03]
J. Wan, Y. Zhou, X. Sun, and M. Zhang. Guaranteeing quality of
service in optical burst switching networks based on dynamic wavelength routing. Optics Communications, 220(1-3):85–95, May 2003.
[WZV02]
J. White, M. Zukerman, and H. L. Vu. A framework for optical burst
switching network design. IEEE Communications Letters, 6(6):268–
270, June 2002.
[XPR01]
L. Xu, H.G. Perros, and G. Rouskas. Techniques for optical packet
switching and optical burst switching. IEEE Communications Magazine, 39(1):136–142, January 2001.
[XQLX03]
J. Xu, C. Qiao, Jikai Li, and G. Xu. Efficient channel scheduling
algorithms in optical burst switched networks. In Proceedings of
IEEE INFOCOM 2003, New York, NY (USA), March-April 2003.
[XTGE+ 04]
Y. Xin, J. Teng, G.Karmous-Edwards, G. Rouskas, and D. Stevenson. Fault management with fast restoration for optical burst
switched networks. In Proceedings of the 1st Annual International
Conference on Broadband Networks (BROADNETS 2004), San Jose,
CA (USA), October 2004.
[XVC99]
Y. Xiong, M. Vandenhoute, and H. Cankaya. Design and analysis
of optical burst-switched networks. In Proceedings of SPIE99 Conference on All Optical Networking, Boston, MA (USA), September
1999.
[XVC00]
Y. Xiong, M. Vanderhoute, and C. Cankaya. Control architecture
in optical burst-switched wdm networks. IEEE Journal of Selected
Areas in Communications, 18(10):1838–1851, October 2000.
Bibliography
146
[YJJ03]
L. Yang, Y. Jiang, and S. Jiang. A probabilistic preemptive scheme
for providing service differentiation in obs networks. In Proceedings of
the IEEE Global Communications Conference (GLOBECOM 2003),
Singapore, December 2003.
[YLC+ 04]
X. Yu, J. Li, X. Cao, Y. Chen, and C. Qiao. Traffic statistics and
performance evaluation in optical burst switched networks. IEEE
Journal of Lightwave Technology, 22(12):2722–2738, December 2004.
[YMY01]
S. Yao, B. Mukherjee, and S.J.B. Yoo. A comparison study between slotted and unslotted all-optical packet-switched network with
priority-based routing. In Proceedings of OFC 2001, Anaheim, CA
(USA), March 2001.
[YQ97]
M. Yoo and C. Qiao. Just-enough-time (jet): a high speed protocol
for bursty traffic in optical networks. In Proceedings of IEEE/LEOS
Technologies for a Global Information Infrastructure (IEEE 1997),
Montreal, Canada, August 1997.
[YQ98]
M. Yoo and C. Qiao. New optical burst switching (obs) protocol
for supporting quality of service. SPIE Proceedings, All Optical Networking: Architecture, Control and Management Issues, 3531:396–
405, November 1998.
[YQD01]
M. Yoo, C. Qiao, and S. Dixit. Optical burst switching for service
differentiation in the next-generation optical internet. IEEE Communications Magazine, 39(2):98–104, February 2001.
[YR06a]
L. Yang and G. N. Rouskas. Adaptive path selection in optical burst
switched networks. IEEE/OSA Journal of Lightwave Technology,
24(8):3002–3011, August 2006.
[YR06b]
Li Yang and George N. Rouskas. A framework for absolute qos
guarantees in optical burst switched networks. In Proceedings of
IEEE Broadnets 2006, San Jose, CA (USA), October 2006.
[ZCC+ 04]
P. Zaffoni, F. Callegati, W. Cerroni, G. Muretto, and C. Raffaelli.
Qos routing in dwdm optical packet networks. In Proceedings of
WQoSR2004 co-located with QoFIS 2004, Barcelona, Spain, September 2004.
[ZJM00]
H. Zang, J. Jue, and B. Mukherjee. A review of routing and wavelength assignment approaches for wavelength-routed optical wdm
networks. Optical Networks Magazine, 1(1):47–60, January 2000.
[ZLW+ 04]
J. Zhang, H. J. Lee, S. Wang, X. Qiu, K. Zhu, Y. Huang, D. Datta,
Y. C. Kim, and B.Mukherjee. Explicit routing for traffic engineering
in labeled optical burst-switched wdm networks. In Proceedings of
Bibliography
147
International Conference on Computational Science (ICCS 2004),
Krakow, Poland, June 2004.
[ZVJC04]
Q. Zhang, V. M. Vokkarane, J. P. Jue, and B. Chen. Absolute
qos differentiation in optical burst-switched networks. IEEE Journal
on Selected Areas in Communications, 22(9):1781–1795, November
2004.
[ZVR+ 04]
A. Zalesky, H.L. Vu, Z. Rosberg, E.W.M. Wong, and M. Zukerman.
Modelling and performance evaluation of optical burst switched networks with deflection routing and wavelength reservation. In Proceedings of the 23rd Joint Conference of IEEE Computer and Communications Societies (INFOCOM 2004), Hong Kong, China, March
2004.
[ZWZ+ 04]
J. Zhang, S. Wang, K. Zhu, D. Datta, Y.-C. Kim, and B. Mukherjee.
Pre-planned global rerouting for fault management in labeled optical burst-switched wdm networks. In Proceedings of Global Telecommunications Conference (GLOBECOM 2004), Dallas, TX (USA),
December 2004.
Appendix A
Related publications
A.1
Papers
1. M. Klinkowski, M. Pioro, D. Careglio, M. Marciniak and J. Sole-Pareta, “Nonlinear Optimization for Multipath Source-Routing in OBS Networks”, IEEE
Communications Letters, vol. 11, no. 12, December 2007.
2. M. Klinkowski, M. Pioro and M. Marciniak, ”Optimization of routing in
optical burst switching networks: a multi-path routing approach”, Chapter of
the COST 293 book (being edited).
3. M. Klinkowski, D. Careglio and J. Solé-Pareta, ”Performance Overview of
QoS Mechanisms for OBS”, Chapter of the book ”Current Research Progress
of Optical Networks” (being edited).
4. M. Klinkowski, D.Careglio and J. Solé-Pareta, “Modelling of Control Plane
in OBS Networks”, in Proceedings of the 9th IEEE International Conference
on Transparent Optical Networks (ICTON2007), Rome, Italy, July 2007.
5. M. Klinkowski, M. Pioro, D.Careglio, M. Marciniak and J. Solé-Pareta, “Routing Optimization in Optical Burst Switching Networks”, in Proceedings of
the 11th Conference on Optical Network Design and Modelling (ONDM2007),
Athens, Greece, May 2007.
6. O. González de Dios, M. Klinkowski, C. Garcı́a Argos, D. Careglio, J. SoléPareta, “Performance Analysis of Routing Algorithms for Optical Burst Switching”, in Proceedings of the 11th Conference on Optical Network Design and
Modelling (ONDM2007), Athens, Greece, May 2007.
7. M. Klinkowski, M. Marciniak, D.Careglio and J. Solé-Pareta, ”Evaluation of
Quality of Service Mechanisms in Optical Burst Switched Networks”, the 4rd
Workshop on Optimization of Optical Networks (OON2007), Montreal, Canada,
May 2007.
149
Appendix A. Related publications
150
8. M. Klinkowski, M. Pioro, D.Careglio, M. Marciniak and J. Solé-Pareta, “Routing Optimization in OBS networks”, COST 293 GRAAL and COST 295 DYNAMO Discussion Workshop, Maribor, Slovenia, January/February 2007.
9. J. Aracil, N. Akar, S. Bjørnstad, M. Casoni, K. Christodoulopoulos, D. Careglio,
J. Fdez-Palacios, C. Gauger, O. Gonzalez de Dios, G. Hu, E. Karasan, M.
Klinkowski, D. Morato, R. Nejabati, H. Øverby, C. Raffaelli, D. Simeonidou,
N. Stol, G. Tosi-Beleffi and K. Vlachos, ”Research in Optical Burst Switching
within the e-Photon/ONe Network of Excellence”, Elsevier Optical Switching
and Networking (OSN) journal, vol. 4, no. 1, pp. 1-19, February 2007.
10. M. Klinkowski, D. Careglio and J. Solé-Pareta, “Offset Time Emulated OBS
Control Architecture”, in Proceedings of the 32nd European Conference on
Optical Communication (ECOC2006), Cannes, France, September 2006.
11. M. Klinkowski, D. Careglio, M. Marciniak and J. Solé-Pareta, “Comparative
Study of QoS Mechanisms in OBS Networks” , in Proceedings of the 11th
European Conference on Networks and Optical Communications (NOC2006),
Berlin, Germany, July 2006.
12. M. Klinkowski, D. Careglio and J. Solé-Pareta, “Comparison of Conventional
and Offset Time-Emulated Optical Burst Switching Architectures”, in Proceedings of the 8th IEEE International Conference on Transparent Optical Networks
(ICTON2006), Nottingham, UK, June 2006.
13. Javier Aracil and S. Bjornstad, M. Casoni, K. Christodoulopoulos, J. FdezPalacios, C. Gauger, O. Gonzalez, G. Hu, E. Karasan, M. Klinkowski, D.
Morato, R. Nejabati, H. Overby, C. Raffaelli, D. Simeonidou, J. Sole-Pareta,
N. Stol, G.M. Tosi Beleffi and K. Vlachos, “The Research Agenda in Optical
Burst Switching in e-Photon/ONe (Tutorial)”, in Proceedings of the 8th IEEE
International Conference on Transparent Optical Networks (ICTON2006), Nottingham, UK, June 2006.
14. M. Klinkowski, D. Careglio, D. Morató and J. Solé-Pareta, “Effective Burst
Preemption in OBS Network”, in Proceedings of IEEE International Workshop
on High Performance Switching and Routing (HPSR 2006), Poznan, Poland,
June 2006.
15. F. Callegati, J. Aracil, L. Wosinska, N. Andriolli, D. Careglio, M. Klinkowski,
A. Giorgetti, J. Fdez-Palacios, C. Gauger, O. Gonzáles de Dios, G. Hu, E.
Karasan, F. Matera, H. Overby, C. Raffaelli, L. Rea, N. Sengezer, M. Tornatore
and K. Vlachos, “Research on Optical Core Networks in the e-Photon/ONe
Network of Excellence”, in Proceedings of IEEE INFOCOM High Speed Networking Workshop: The Terabits Challenge (co-located with INFOCOM2006),
Barcelona, Spain, April 2006.
16. E. Hortas, D. Careglio, M. Klinkowski and J. Solé-Pareta, “Análisis de Prestaciones de Algoritmos de Encaminamiento Adaptativos para la Conmutación de
Appendix A. Related publications
151
Ráfagas Ópticas”, XV Jornadas Telecom I+D, Madrid / Barcelona / Valencia,
Spain, November 2005.
17. M. Klinkowski, D. Careglio, E. Hortas and J. Solé-Pareta, “Performance
Analysis of Isolated Adaptive Routing Algorithms in OBS Networks”, in Proceedings of e-Photon/ONe Summer School workshop, Rimini, Italy, August
2005.
18. M. Klinkowski, D. Careglio, S. Spadaro and J. Solé-Pareta, “Impact of Burst
Length Differentiation on QoS Performance in OBS Networks”, in Proceedings
of the 7th IEEE International Conference on Transparent Optical Networks
(ICTON2005), Barcelona, Spain, July 2005.
19. M. Klinkowski, D. Careglio and J. Solé-Pareta, “Wavelength vs Burst vs
Packet Switching: Comparison of Optical Network Models”, in Proceedings of
e-Photon/ONe Winter School workshop, Aveiro, Portugal, February 2005.
20. M. Klinkowski, F. Herrero, D. Careglio and J. Solé-Pareta, ”Adaptive Routing
Algorithms for Optical Packet Switching Networks”, in Proceedings of 9th IFIP
Working Conference on Optical Network Design and Modelling (ONDM2005),
Milan, Italy, February 2005.
21. F. Herrero, D. Careglio, J. Solé-Pareta and M. Klinkowski, “Algoritmos de
Enrutamiento para la Conmutación de Paquetes Ópticos”, XIV Jornadas Telecom I+D, Madrid/Barcelona/Valencia, November 2004.
22. M. Klinkowski, D. Careglio, X. Masip-Bruin, S. Spadaro, S. Sanchez-López
and J. Solé-Pareta, “A Simulation Study of Combined Routing and Contention
Resolution Algorithms in Connection-Oriented OPS Network Scenario”, in Proceedings of the 6th IEEE International Conference on Transparent Optical Networks (ICTON2004), Wroclaw, Poland, July 2004.
23. S. Bjornstad, C. M. Gauger, M. Nord, E. Baert, F. Callegati, D. Careglio,
M. Klinkowski, M. Marciniak, J. Solé-Pareta et al, “Optical Burst Switching
and Optical Packet Switching”. Chapter 4 of Book Advanced Infrastructure for
Photonic Networks – Extended Final Report of COST Action 266, pp. 115-154.
R. Inkret et al., Editorial Faculty of Electrical, Engineering and Computing,
University of Zagreb, 2003, ISBN 953-184-064-4.
24. S. Bjornstad, M. Nord, D.R. Hjelme, N. Stol, F. Callegati, W. Cerroni, C.
Raffaelli, P. Zaffoni, C. M. Gauger, C. Develder, J. Cheyns, E. Van Breusegem,
E. Baert, D. Colle, M. Pickavet, P. Demeester, M. Lackovic, D. Careglio, G.
Junyent, M. Klinkowski, M. Marciniak and M. Kowalewski, “Optical Burst
and Packet Switching: Node and Network Design, Contention Resolution and
Quality of Service”, in Proceedings of the 7th IEEE International Conference
on Telecommunications (ConTEL2003), Zagreb, Croatia, June 2003.
Appendix A. Related publications
A.2
152
Papers under submission
1. M. Klinkowski, D. Careglio and J. Solé-Pareta, “Reactive and Proactive Routing in Labelled OBS Networks”, submitted to IET Journal.
2. M. Klinkowski, D. Careglio, Daniel Morató and J. Solé-Pareta, “Preemption
Window for Burst Differentiation in OBS”, submitted to OFC 2008 Conference.
A.3
Contribution to European projects
1. Deliverable D3.1, “Architectures and preliminary definition of specific CP and
MP functions for hybrid opto-electronic burst/packet networks”, FP6-027305
IP Nobel 2 Project, October 2006.
2. Deliverable D32, “Preliminary report on feasibility studies on opto-electronic
burst/packet switching nodes”, FP6-506760 IP Nobel Project, January 2006.
3. Deliverable D23, “Definition of hybrid opto-electronic burst/packet switching
node structures and related management functions”, FP6-506760 IP Nobel
Project, September 2005.
4. Deliverable D16, “Preliminary definition of burst/packet network and node architectures and solutions”, FP6-506760 IP Nobel Project, March 2005.
5. M. Klinkowski, M. Marciniak et al, “Optical Packet Router with QoS Capabilities: Introductory Study of Computer Simulator Design”. In Progress Report
of COST 266 Action, June 2002.
A.4
Other publications
1. P. Pedroso, J. Solé-Pareta, D. Careglio and M. Klinkowski, “Integrating GMPLS in the OBS Networks Control Plane”, in Proceedings of the 9th IEEE International Conference on Transparent Optical Networks (ICTON2007), Rome,
Italy, July 2007.
2. E. Bonada, F. Callegati, D. Careglio, W. Cerroni, M. Klinkowski, G. Muretto,
C. Raffaelli and J. Solé-Pareta, “SCWS Technique for QoS Support in ConnectionOriented Optical Packet Switching Network”, in Proceedings of the 8th IEEE
International Conference on Transparent Optical Networks (ICTON2006), Nottingham, UK, June 2006.
3. D. Careglio, J. Solé-Pareta, M. Klinkowski and S. Spadaro, “Modelling and
Optimisation of the IST DAVID Metro Networks”, in Proceedings of the 10th
European Conference on Networks and Optical Communications (NOC2005),
invited paper, London, UK, July 2005.
Appendix A. Related publications
153
4. M. Klinkowski, D. Careglio, M. Marciniak and J. Solé-Pareta, “Performance
Analysis of the Simple Prioritized Buffering Algorithm in Optical Packet Switch
for DiffServ Assured Forwarding”, in Proceedings of the 5th IEEE International
Conference on Transparent Optical Networks (ICTON2003), Warsaw, Poland,
June 2003.
5. M. Klinkowski and M. Marciniak, “Services Differentiation in MPLS Photonic
Packet Networks”, in Proceedings of the 7th IFIP Working Conference on Optical Network Design and Modeling (ONDM2003) - COST266 IST OPTIMIST
Workshop, Budapest, Hungary, February 2003.
6. M. Klinkowski and M. Marciniak, “Optical Packet Switching: Approach to
Performance Modeling and Simulation”, in Proceedings of the 27th triennial
General Assembly of the International Union of Radio Science (URSI2002),
Maastricht (NL), August 2002.
7. M.Marciniak and M.Klinkowski, “Advanced Optical Infrastructure for the
Emerging Optical Internet Services”, in Proceedings of the 3th International
Conference on Advances in Infrastructure for e-Business, e-Education, e-Science,
and e-Medicine on the Internet (SSGRR2002), L’Aquila, Italy, July/August
2002.
8. M.Klinkowski and M.Marciniak, “Optical Packet Router with QoS Capabilities Introductory Study of Computer Simulator Design”, in Proceedings of
the 4th IEEE International Conference on Transparent Optical Networks (ICTON2002), Warsaw, Poland, April 2002.
9. M. Klinkowski and M. Marciniak, “QoS Guarantees in IP Optical Networks
using MPLS/ MPLambdaS”, in Proceedings of the 3rd IEEE International Conference on Transparent Optical Networks (ICTON2001), Cracow, Poland, June
2001.
10. M. Klinkowski and M. Marciniak, “Development of IP/WDM Optical Networks”, in Proceedings of the 3th International Workshop on Laser and Fiber
optic Network Modeling (LFNM2001), Kharkov, Ukraine, May 2001.
11. M. Klinkowski and M. Marciniak, “IP over Optical Network: Strategy of Deployment”, Journal of Telecommunications and Information Technology, NIT,
no. 2, pp. 51-56, April 2001.